feat: add LLM orchestrator, blue-team investigation pipeline, and benchmark replay system by l50 · Pull Request #441 · dreadnode/ares

l50 · 2026-08-10T04:24:27Z

Key Changes:

Reintroduced an LLM-driven orchestrator with proposal-based work mediation, configurable reasoning effort, per-role token/cost accounting, and hard wall-clock timeouts with self-healing
Added a comprehensive blue-team investigation pipeline with detection sweeps, simulated containment loops, coverage measurement against red ground truth, and Redis-backed escalation state
Built an end-to-end benchmark capture/replay system with Tempo trace capture, S3 snapshots, and a self-contained Grafana/Loki/Mimir replay stack
Migrated Ansible attack tooling and cloud/monitoring roles to external l50.arsenal and l50.bulwark collections, consolidating the local role tree

Added:

LLM orchestrator subsystem - New orchestrator/proposals.rs, flags.rs, orchestrator_planning.rs, and planner prompt templates enabling proposal-based mediation, config-backed flags, discovery-driven wakes, and advancement-based stall detection
Blue-team investigation and reporting - Added orchestrator/blue/sweep.rs, simulated_response.rs, cleanup/ engine (baseline, capture, journal, registry), blue_invalidation.rs, and blue coverage/provenance report modules for containment tracking and ground-truth coverage
Benchmark replay pipeline - New ares-cli/src/benchmark/ (capture, replay, manifest, tempo_push, snapshot_s3), benchmarks/replay-stack/ compose stack, and ares-replay-stack warpgate template plus docs/benchmark-replay.md
ADCS/ACL attack graph tooling - Added orchestrator/acl_graph.rs, BloodHound parser (parsers/bloodhound.rs), certipy shadow-credential and ESC1/ESC3/ESC13/ESC16 chain support, and owner_edit/windows_stage_and_run tools
New credential and lateral tooling - Added local_auth_sweep.rs, mssql_link_pivot.rs, credential_reuse.rs, silver/golden ticket forging with correlation, and windows_payload.rs privilege-escalation staging
Infrastructure and skills - Added .claude/.gemini skill suites, golden AMI build scripts, vector and sysmon Ansible roles, k8s/obs/proxmox Taskfiles, redaction/sanitization utilities, and demo/preflight.sh
Claude CLI provider - New ares-llm/src/provider/claude_cli.rs and reasoning-effort configuration for reasoning models

Changed:

Trust and delegation automation - Substantially expanded automation/trust.rs, dacl_abuse.rs, s4u.rs, and secretsdump.rs to harden cross-forest forges, credential attribution, and Kerberos ticket path reconciliation
Result processing and state - Reworked result_processing/mod.rs, state/inner.rs, and publishing modules to add containment recovery, ACL grant handling, credential provenance attestation, and domain canonicalization
Detection and MITRE mapping - Expanded detections.yaml and mitre_techniques.yaml with ATT&CK path mappings, shadow-credential/ACL/NoPac detections, and catalog-grounded technique enforcement
Crack task dispatch - Hardened automation/crack.rs and deferred queue to dedupe by hash identity and reserve dedup keys at submission to close double-queue races
Config consolidation - Established config/ares.yaml as single source of truth for model selection, added base-URL override and cached-token accounting, and removed legacy Grafana credential fields
Taskfile and EC2 automation - Rewrote .taskfiles/ec2/, blue/, red/, and remote/ Taskfiles with reusable SSM/EC2/k8s shell helpers and an e2e-op.sh end-to-end runner
Documentation - Consolidated blue/red team docs, updated infrastructure references, and migrated template docs to the main branch

Removed:

Local Ansible attack roles - Removed acl_tools, coercion_tools, cracking_tools, credential_access_tools, lateral_movement_tools, privesc_tools, and recon_tools roles in favor of external collections
Local cloud/monitoring roles - Removed alloy, aws_cloudwatch_agent, aws_ssm_agent, dc_audit_sacl, and mythic roles now sourced from l50.bulwark
KrbRelayUp automation - Removed automation/krbrelayup.rs and the remote crackd hashcat backend
Legacy orchestrator and docs - Retired the prior orchestrator agent/trap dispatch tools, SeImpersonate escalation automation, docs/plan-loot-gaps.md, and stale Grafana MCP docs

**Key Changes:** - Added retry loop for transient `KRB_AP_ERR_MODIFIED` PKINIT failures during certipy auth, retrying up to 4 times - Fixed auth command to pass bare sAMAccountName via `-username` instead of UPN form, preventing AS-REP failures for RID-500 Administrator - Corrected overall success determination to handle RC4-disabled KDCs where certipy auth exits non-zero despite producing a valid ccache **Changed:** - certipy auth invocation - Replaced single-shot `certipy auth` command with a retry loop (up to 4 attempts) that detects transient `KRB_AP_ERR_MODIFIED` errors caused by intermittent DH/session-key mismatches on AES-only KDCs; also added `-username` flag passing the bare sAMAccountName (split from UPN) to prevent PKINIT AS-REP failures for the built-in Administrator account (`certipy_esc1_full_chain` in `adcs.rs`) - Success and exit code logic - Replaced the previous success check (`request_output.success && auth_output.success && dcsync_success`) with a more nuanced `overall_success` determination: when DCSync ran, both the request and DCSync steps must succeed (auth exit code is not authoritative on RC4-disabled KDCs); when no DCSync tail ran, success additionally requires `got_nt_hash` to ensure a clean auth that recovered no hash is reported as a failure rather than silently deduped as complete (`certipy_esc1_full_chain` in `adcs.rs`)

…n ticket stop condition (#242) **Key Changes:** - Added `--potfile-disable` to every hashcat invocation to prevent persistent plaintext accumulation on disk - Enabled `stop_on_golden_ticket` in the default operation config to halt on golden ticket acquisition - Fixed a shell printf format string in the EC2 Taskfile to avoid potential flag misinterpretation **Added:** - Potfile suppression across all hashcat passes - Added `--potfile-disable` argument in `niced_hashcat()` so no cross-operation state accumulates on disk; cracks are recovered from per-pass stdout instead, ensuring no plaintext is lost (`ares-tools/src/cracker.rs`) - Doc comment explaining potfile rationale - Added inline documentation to `niced_hashcat()` describing why the potfile is disabled and how crack recovery works without it - Test assertion for potfile flag - Extended the existing workload profile test to verify `--potfile-disable` is present in every hashcat command's argument list **Changed:** - Golden ticket stop condition - Flipped `stop_on_golden_ticket` from `false` to `true` in `config/ares.yaml`, making operation halt automatically upon golden ticket acquisition rather than continuing - EC2 profile flag formatting - Corrected `printf -- '--profile %s'` to `printf '%s' '--profile {{.AWS_PROFILE}}'` in `.taskfiles/ec2/Taskfile.yaml` to properly pass the profile flag string and avoid shells interpreting it as a printf option

**Key Changes:** - Fixed a bug where canonical hyphenated hash type spellings (e.g., "AS-REP", "TGS-REP") were misclassified as priority-1 instead of priority-0, causing crackable Kerberos tickets to be starved behind NTLM floods - Introduced punctuation stripping before match comparison so that normalized forms from `dedup::normalize_hash_type` correctly collapse onto bare roast tokens - Expanded the priority-0 match arm to include additional Kerberos ticket spellings (`krb5asrep`, `krb5tgs`, `tgsrep`, `tgs`) - Added a regression test covering all affected canonical and variant spellings **Added:** - Regression test for canonical roast spelling normalization - `crack_priority_normalizes_canonical_roast_spellings` verifies that both hyphenated canonical forms and bare token variants are correctly ranked as priority-0, and that non-roastable types like NTLM remain priority-1 **Changed:** - Hash type normalization in priority ranking - `crack_priority` in `crack.rs` now strips `-` and `_` characters after lowercasing before matching, mirroring the behavior of `credential_resolver::is_authenticating_hash_type`; previously "AS-REP" lowercased to "as-rep" which never matched the "asrep" arm, misclassifying roastable tickets as priority-1 and allowing the secretsdump NTLM flood to starve genuinely crackable AS-REP and TGS-REP tickets indefinitely

…ed cross-forest admin (#244) **Key Changes:** - Introduced `is_native_adcs_enum_candidate` to identify same-domain, non-machine accounts with recovered passwords as preferred ADCS enumerators - ADCS enumeration now selects a native domain user when available, ensuring certipy correctly flags ESC1 (enrollee-supplies-subject) templates as vulnerable - Fell back to the forged cross-forest Administrator only when no native credential exists, preserving existing behavior for uncracked environments - Added unit tests covering all candidate selection edge cases **Added:** - Native enrollee selection logic - `is_native_adcs_enum_candidate` predicate filters for same-domain, non-machine accounts with a non-empty plaintext password; delegation accounts and quarantined principals are excluded by the caller in `dispatch_post_ticket_adcs_enumeration` - Unit test suite for candidate selection - `native_adcs_enum_candidate_prefers_same_domain_password_user` covers eligible native user, case-insensitive domain match, cross-forest account exclusion, machine account (`$` suffix) exclusion, and uncracked credential exclusion **Changed:** - ADCS enumeration principal selection - `dispatch_post_ticket_adcs_enumeration` now resolves a `native_user` from state alongside `target_dc_ip` and passes it as the `username` in `tool_args`, falling back to `"Administrator"` only when no native candidate is found; this ensures certipy binds as a Domain Users member so ESC1 templates are surfaced rather than silently skipped - Structured log output for ADCS dispatch - added `principal` and `native` fields to the tracing event so operators can confirm which account was used for each enumeration run

| datasource | package | from | to | | ----------- | -------------------- | ------- | ------- | | github-tags | github/codeql-action | v4.37.1 | v4.37.2 |

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [taiki-e/install-action](https://redirect.github.com/taiki-e/install-action) ([changelog](https://redirect.github.com/taiki-e/install-action/compare/07b4745e0c39a41822af610387492e3e53aa222b..a6b2e2dcd845ddd7f509ce4f3ed3d922b80cc5d9)) | action | digest | `07b4745` → `a6b2e2d` | --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Enabled. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

| datasource | package | from | to | | ----------- | ---------------- | ------ | ------ | | github-tags | actions/checkout | v7.0.0 | v7.0.1 |

| datasource | package | from | to | | ----------- | ------------------------- | -------- | -------- | | github-tags | renovatebot/github-action | v46.1.19 | v46.1.20 |

| datasource | package | from | to | | ----------- | --------------- | ------ | ------ | | github-tags | actions/labeler | v6.2.0 | v7.0.0 |

| datasource | package | from | to | | ----------- | -------------------- | ------ | ------ | | github-tags | actions/setup-python | v6.3.0 | v7.0.0 |

This PR contains the following updates: | Package | Change | [Age](https://docs.renovatebot.com/merge-confidence/) | [Confidence](https://docs.renovatebot.com/merge-confidence/) | |---|---|---|---| | [pre-commit](https://redirect.github.com/pre-commit/pre-commit) | `==4.6.0` → `==4.6.1` | ![age](https://developer.mend.io/api/mc/badges/age/pypi/pre-commit/4.6.1?slim=true) | ![confidence](https://developer.mend.io/api/mc/badges/confidence/pypi/pre-commit/4.6.0/4.6.1?slim=true) | --- ### Release Notes <details> <summary>pre-commit/pre-commit (pre-commit)</summary> ### [`v4.6.1`](https://redirect.github.com/pre-commit/pre-commit/blob/HEAD/CHANGELOG.md#461---2026-07-21) [Compare Source](https://redirect.github.com/pre-commit/pre-commit/compare/v4.6.0...v4.6.1) \================== ##### Fixes - Install `language: node` hooks via `git`. - Fixes npm 12.x compatibility - [#3719](https://redirect.github.com/pre-commit/pre-commit/issues/3719) PR by [@asottile](https://redirect.github.com/asottile). - [#3517](https://redirect.github.com/pre-commit/pre-commit/issues/3517) issue by [@ojob](https://redirect.github.com/ojob). - Set `JULIA_DEPOT_PATH` for `language: julia`. - [#3711](https://redirect.github.com/pre-commit/pre-commit/issues/3711) PR by [@damonbayer](https://redirect.github.com/damonbayer). - [pre-commit-ci/runner-image#335](https://redirect.github.com/pre-commit-ci/runner-image/issues/335) issue by [@damonbayer](https://redirect.github.com/damonbayer). - Produce error on mistyped `--repo` for `pre-commit autoupdate`. - [#3701](https://redirect.github.com/pre-commit/pre-commit/issues/3701) PR by [@mxr](https://redirect.github.com/mxr). - [#3695](https://redirect.github.com/pre-commit/pre-commit/issues/3695) issue by [@mxr](https://redirect.github.com/mxr). - Improve performance of commit existence check in `pre-push`. - [#3726](https://redirect.github.com/pre-commit/pre-commit/issues/3726) PR by [@asottile](https://redirect.github.com/asottile). - [#3604](https://redirect.github.com/pre-commit/pre-commit/issues/3604) issue by [@ptarjan](https://redirect.github.com/ptarjan). - Avoid duplicating conflicted filenames during `pre-commit run --all-files`. - [#3727](https://redirect.github.com/pre-commit/pre-commit/issues/3727) PR by [@asottile](https://redirect.github.com/asottile). - [#3706](https://redirect.github.com/pre-commit/pre-commit/issues/3706) issue by [@RomanValov](https://redirect.github.com/RomanValov). </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Update | Change | |---|---|---| | [grafana/grafana](https://redirect.github.com/grafana/grafana) | patch | `13.1.0` → `13.1.1` | --- ### Release Notes <details> <summary>grafana/grafana (grafana/grafana)</summary> ### [`v13.1.1`](https://redirect.github.com/grafana/grafana/releases/tag/v13.1.1): 13.1.1 [Download page](https://grafana.com/grafana/download/13.1.1) [What's new highlights](https://grafana.com/docs/grafana/latest/whatsnew/) ##### Features and enhancements - **Go:** Update version to 1.26.5 [#128015](https://redirect.github.com/grafana/grafana/pull/128015), [@macabu](https://redirect.github.com/macabu) - **Provisioning:** Improve form errors for github connections [#128177](https://redirect.github.com/grafana/grafana/pull/128177), [@grafana-writer\[bot\]](https://redirect.github.com/grafana-writer\[bot]) - **Provisioning:** make sync per-resource write timeout configurable [#127868](https://redirect.github.com/grafana/grafana/pull/127868), [@grafana-writer\[bot\]](https://redirect.github.com/grafana-writer\[bot]) ##### Bug fixes - **Accessibility:** Ensure `InlineToast` contents are announced by screenreaders [#128687](https://redirect.github.com/grafana/grafana/pull/128687), [@grafana-writer\[bot\]](https://redirect.github.com/grafana-writer\[bot]) - **DashboardDS:** Fix chained dashboard datasource panels showing stale data [#127248](https://redirect.github.com/grafana/grafana/pull/127248), [@grafana-writer\[bot\]](https://redirect.github.com/grafana-writer\[bot]) - **Provisioning:** make GitHub webhook creation idempotent (fix repos stuck unhealthy with HTTP 422) [#128201](https://redirect.github.com/grafana/grafana/pull/128201), [@floriecai](https://redirect.github.com/floriecai) ##### Plugin development fixes & changes - **Pagination:** Set `aria-current` on active page [#128518](https://redirect.github.com/grafana/grafana/pull/128518), [@grafana-writer\[bot\]](https://redirect.github.com/grafana-writer\[bot]) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [serde_json](https://redirect.github.com/serde-rs/json) | workspace.dependencies | patch | `1.0.150` → `1.0.151` | --- ### Release Notes <details> <summary>serde-rs/json (serde_json)</summary> ### [`v1.0.151`](https://redirect.github.com/serde-rs/json/releases/tag/v1.0.151) [Compare Source](https://redirect.github.com/serde-rs/json/compare/v1.0.150...v1.0.151) - Add RawValue::from\_string\_unchecked ([#1331](https://redirect.github.com/serde-rs/json/issues/1331), thanks [@WonderLawrence](https://redirect.github.com/WonderLawrence)) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [tokio](https://tokio.rs) ([source](https://redirect.github.com/tokio-rs/tokio)) | workspace.dependencies | patch | `1.53.0` → `1.53.1` | --- ### Release Notes <details> <summary>tokio-rs/tokio (tokio)</summary> ### [`v1.53.1`](https://redirect.github.com/tokio-rs/tokio/releases/tag/tokio-1.53.1): Tokio v1.53.1 [Compare Source](https://redirect.github.com/tokio-rs/tokio/compare/tokio-1.53.0...tokio-1.53.1) ### 1.53.1 (July 20th, 2026) ##### Fixed - signal: restore MSRV by removing `OnceLock::wait` from the Windows handler ([#8300]) ##### Fixed (unstable) - time: fix alt timer cancellation and insertion race ([#8252]) ##### Documented - runtime: remove dead link definition in Runtime::block\_on ([#8301]) [#8252]: https://redirect.github.com/tokio-rs/tokio/pull/8252 [#8300]: https://redirect.github.com/tokio-rs/tokio/pull/8300 [#8301]: https://redirect.github.com/tokio-rs/tokio/pull/8301 </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [async-nats](https://redirect.github.com/nats-io/nats.rs) | workspace.dependencies | minor | `0.49` → `0.50` | --- ### Release Notes <details> <summary>nats-io/nats.rs (async-nats)</summary> ### [`v0.50.0`](https://redirect.github.com/nats-io/nats.rs/releases/tag/async-nats/v0.50.0) #### Overview This release allows for on-demand swap between `chrono` and `time` crates. #### What's Changed - Add chrono as alternative to time crate by [@Jarema](https://redirect.github.com/Jarema) in [#1595](https://redirect.github.com/nats-io/nats.rs/pull/1595) - Use new start method for retry start in nats-server crate by [@Jarema](https://redirect.github.com/Jarema) in [#1603](https://redirect.github.com/nats-io/nats.rs/pull/1603) - Fix account info deser failure on servers with tiered jetstream by [@xanderio](https://redirect.github.com/xanderio) in [#1604](https://redirect.github.com/nats-io/nats.rs/pull/1604) #### Chrono vs Time Enabling `chrono` anywhere in the dependency graph selects the chrono backend for the whole build (Cargo feature unification). #### New Contributors - [@xanderio](https://redirect.github.com/xanderio) made their first contribution in [#1604](https://redirect.github.com/nats-io/nats.rs/pull/1604) **Full Changelog**: <nats-io/nats.rs@async-nats/v0.49.1...async-nats/v0.50.0> </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

…an-mark timing fix (#256) **Key Changes:** - Implemented full ESC13 (issuance-policy OID → group link) exploitation chain with correct plain-enrollment semantics — no `-upn`/`-sid` override — preventing CA policy module rejections and KB5014754 strict-SID mapping failures - Fixed scan target deduplication mark timing: moved `scanned_targets` registration from submit-request time to actual-dispatch time, preventing deferred/evicted scan tasks from permanently suppressing host port scans - Enabled gMSA password retrieval via machine-account NT hash (`-H` auth), allowing the gMSA's sole authorized reader (`HOST$`) to be used when it has no plaintext credential - Fixed RID-less `$MACHINE.ACC` secretsdump row parsing so the dumped host's own machine account is captured rather than silently dropped **Added:** - ESC13 full chain tool (`certipy_esc13_full_chain`) — plain enrollment, PKINIT auth, and DCSync `krbtgt` with the elevated ccache; registered in `ares-tools/src/lib.rs` and `parsers/mod.rs` alongside `certipy_esc1_full_chain` - `dispatch_esc13_deterministic` and `build_esc13_chain_args` — deterministic ESC13 dispatcher with the same dedup/retry lifecycle as ESC1, explicitly excluding `-upn`/`-sid` args; routing added to `auto_adcs_exploitation` before the ESC8 branch - `is_owned_domain_ntlm` filter in `crack.rs` — skips NTLM hashes from already-dominated domains to prevent them from starving AS-REP/kerberoast crack slots for forests not yet owned - `scan_target_from_payload` helper in `submission.rs` — extracts the target IP from recon payloads carrying `network_scan` or `nmap_scan` techniques, used to gate the now-deferred scan mark - `RE_NTLM_MACHINE_ACCT` regex in `hashes.rs` — dedicated pass for RID-less `DOMAIN\HOST$:LM:NT:::` rows that the existing domain/plain NTLM regexes (which require a numeric RID) would miss - `on_behalf_nt_domain` helper in `adcs.rs` — derives the NetBIOS flat name from the FQDN's first DNS label (uppercased) for certipy's `-on-behalf-of`, fixing ESC3 failures caused by passing an FQDN there - `run_post_ticket_adcs_enum` / `AdcsEnumOutcome` in `trust.rs` — extracted reusable ADCS enumeration primitive that returns a typed outcome; `dispatch_post_ticket_adcs_enumeration` now tries the native enrollee first and falls back to the forged Administrator only when the native run found nothing - `reader_hash: Option<String>` field on `GmsaWork` and supporting resolution logic in `select_gmsa_work` — resolves machine-account readers from the hash store when no plaintext credential exists; payload builder emits `credential.hash` when set - `is_hash32` helper in `secrets.rs` — distinguishes a 32-hex hash from a numeric RID to detect the RID-less `$MACHINE.ACC` shape in secretsdump output **Changed:** - `crack_priority` match arms narrowed to `kerberoast | asrep | asreproast` only — removed `krb5asrep`, `krb5tgs`, `tgsrep`, `tgs` aliases that were never emitted by `normalize_hash_type`, and updated tests to reflect the actual canonical spellings (`AS-REP`, `Kerberoast`) - `parse_secretsdump` in `secrets.rs` — detects RID-less rows (field 1 is a 32-hex hash, not a numeric RID) and reads LM/NT from the shifted positions; RID-less rows are never classified as trust keys since they represent the host's own account - `gmsa_dump_passwords` tool in `privesc/gmsa.rs` — passes the `hash` argument through to `netexec_creds` for `-H` auth; updated tool definition description and schema to document the `hash` field - `submit_to_llm` in `submission.rs` — marks `DEDUP_SCANNED_TARGETS` at actual dispatch time (after throttle and per-credential gates) instead of at submit-request time; removed the early mark from `task_builders.rs` - `dispatch_post_ticket_adcs_enumeration` in `trust.rs` — refactored to try native enrollee first, then fall back to forged Administrator if the native run found zero vulnerabilities, ensuring a stale/rotated native password never leaves the CA un-enumerated - ESC3 `-on-behalf-of` argument in `adcs.rs` — now uses the NetBIOS flat name (`CONTOSO`) instead of the FQDN (`contoso.local`), fixing CA policy module denials (`0x80070547`) - `config/ares.yaml` completion mode defaults — both `stop_on_domain_admin` and `stop_on_golden_ticket` set to `false` (full-forest compromise mode); expanded inline comments document all three stop conditions and the multi-domain caveat for `stop_on_golden_ticket`

…istory DB (#257) **Key Changes:** - Introduced a full mutation journal and teardown system that records every persistent target change an operation makes and can reverse them LIFO via `ares ops teardown`, surviving SIGKILL and post-op worker shutdown - Added pre-op attacker workspace sanitization (hashcat potfile, netexec `~/.nxc` DBs/artifacts, Kerberos ccaches) to prevent ops from inheriting prior op residue, exposed as both an automatic orchestrator pass and `ares ops sanitize` - Replaced the remote RDS dependency with a box-local loopback Postgres for `ares-history`, eliminating cross-region networking for the us-east-1 kali-ares box **Added:** - Mutation journal infrastructure (`orchestrator/cleanup/journal.rs`) — a Redis LIST (`ares:op:{id}:mutation_journal`) that records every successful mutating tool call with target, principal, forward args, and a forward-time hint (e.g. pywhisker DeviceID, noPac machine account name); rides the same 24h TTL as all other op keys so teardown works long after the orchestrator exits - `JournalingToolDispatcher` decorator (`orchestrator/cleanup/dispatcher.rs`) — wraps the single shared `Arc<dyn ToolDispatcher>` so both LLM-driven and deterministic tool calls are captured at one choke point with zero changes to automation modules - Teardown engine (`orchestrator/cleanup/engine.rs`) — reads the journal LIFO, dispatches inverses in-process via `ares_tools::dispatch`, runs optional read-back validation probes, and emits a structured report with per-entry status (Verified / Reverted / Unverified / Skipped / Failed) - Undo registry (`orchestrator/cleanup/registry.rs`) — maps 18 mutating tools to their inverse plans and reversibility classes (CLEAN / NEEDS-CAPTURE / HARD / IMPOSSIBLE / UNSUPPORTED), with read-back probes for `rbcd_write`, `add_computer`, `bloodyad_add_group_member`, and `nopac` - Forward-time state capture (`orchestrator/cleanup/capture.rs`) — scrapes pywhisker DeviceID and noPac machine account name from tool stdout so teardown can build faithful inverses without a read-before-write - `ares ops teardown` subcommand (`ares-cli/src/ops/teardown.rs`) — standalone CLI entry point supporting `--latest`, `--dry-run`, and `--only <tool>`; exits non-zero when any attempted revert fails so `task ec2:teardown` and CI can gate on it - `ares ops sanitize` subcommand (`ares-cli/src/ops/sanitize.rs`) — manually triggers the workspace sanitizer and prints counts for potfile reset, nxc paths removed, and ccaches removed - `ares_tools::sanitize` module (`ares-tools/src/sanitize.rs`) — wipes hashcat potfile, netexec `~/.nxc` workspaces/DBs/spider/artifacts, and `/tmp/ares-tickets` ccaches; respects `ARES_KEEP_WORKSPACE=1`; warns when remote crackd is configured (its server-side potfile is out of reach) - `bloodyad_get_object` tool (`ares-tools/src/acl.rs`) — LDAP read-back used by teardown validation probes to confirm mutations are gone after revert - `task ec2:history-db` and `setup-history-db.sh` — idempotent provisioner for a box-local PostgreSQL `ares_history` database with passwordless loopback trust for `ares_admin`; ares self-migrates the schema on first start - `task ec2:teardown` — task wrapper that invokes `ares ops teardown` over SSM with support for `OPERATION_ID`, `LATEST`, `DRY_RUN`, and `ONLY` parameters **Changed:** - Orchestrator startup (`orchestrator/mod.rs`) — wraps the tool dispatcher with `JournalingToolDispatcher` immediately after construction, and runs the workspace sanitizer before any automation dispatches a tool; also adds an explicit pre-launch `ares ops sanitize` call in the EC2 launch script (belt-and-suspenders alongside the in-binary pass) - Default `ARES_DATABASE_URL` (`Taskfile.yaml`) — changed from empty (RDS-dependent) to `postgresql://ares_admin@127.0.0.1:5432/ares_history` (box-local loopback); DB reachability probe now derives host:port from the URL rather than hardcoding the RDS endpoint, so it works for both local and remote targets - `task ec2:start` — extended to start `postgresql` alongside Redis and NATS; the start is best-effort (no-op if `task ec2:history-db` has not yet run) and the launcher probes 5432 before passing `ARES_DATABASE_URL` to the orchestrator - `add_computer` tool (`ares-tools/src/privesc/delegation.rs`) — added optional `action` parameter (`add` [default] | `delete`) so teardown can delete machine accounts created by an op without a separate tool - `bloodyad_add_group_member` and `bloodyad_add_genericall` (`ares-tools/src/acl.rs`) — added optional `action` parameter (default `add`) so teardown can pass `remove` to reverse group membership and GenericAll grants - `pywhisker` builder (`ares-tools/src/acl.rs`) — added optional `--device-id` flag forwarding so teardown can target the exact Key Credential entry to remove using the captured DeviceID hint - `task ec2:poll-op` precondition — added an explicit `test -x {{.ARES_CLI}}` guard so a missing binary fails fast rather than silently looping until `MAX_WAIT`

…on (#258) **Key Changes:** - Introduced a `simulated_response` module that translates blue escalation decisions (`confirm_escalation`, `downgrade_escalation`) into tracing spans for the demo dashboard and optionally publishes matching op-state events to the red-side projector via NATS - Added `OpStateRecorder` to `BlueCallbackHandler` so confirmed containment actions (disable AD account, isolate host, revoke krbtgt/certificate) invalidate red's in-flight and deferred task queues - Fixed a multi-forest completion bug where a discovered-but-never-exploited trust forge whose target forest was already dominated via another path (native ADCS/DCSync) pinned the op open to the hard max-runtime cap - Added red dispatch freezing at the moment of op completion so the swarm stops burning LLM tokens on the exploit/ACL backlog during the post-completion blue-drain window **Added:** - `simulated_response` module (`ares-cli/src/orchestrator/blue/simulated_response.rs`) — emits `blue.simulated_response.<action_type>` spans (Tempo spanmetrics, demo dashboard `Simulated Response Actions` panel) and provides `payload_for_containment` / `publish_containment` helpers that translate action slugs into `OpStateEventPayload` variants (`CredentialRevoked`, `HostIsolated`, `KrbtgtRotated`, `CertificateRevoked`) - `BlueCallbackHandler::with_recorder` constructor — wires a caller-supplied `OpStateRecorder` (NATS-backed in production, capturing in tests); the recorder-less `new` is preserved for call sites without a broker - `handle_confirm_escalation` and `handle_downgrade_escalation` async handlers — replace the previous static path in `dispatch_callback`, emit spans unconditionally, and publish containment observations when a concrete action + target are named - `task_dropped_by_containment` filter in `deferred.rs` — checks isolated hosts, revoked credentials, and rotated krbtgt against deferred task payloads before re-dispatch, preventing noisy `STATUS_LOGON_FAILURE` / `STATUS_HOST_UNREACHABLE` errors during the demo - `dominated_forest_roots` extracted helper and `escalation_target_forest_dominated` guard in `completion.rs` — shared by `compute_undominated_forests` and `has_pending_cross_forest_escalation` so the two completion guards can't drift - `containment_action` and `target` fields added to the `confirm_escalation` tool schema in `ares-llm` — the LLM can now name a specific containment action and its subject; `escalate_to_human` remains the default no-op **Changed:** - `run_investigation` signature gains an `op_state_recorder` parameter and splices `operation_id` into `alert.labels` before constructing the callback handler, so blue-side spans carry `attack_operation_id` and appear in per-op dashboard filters - `BlueOrchestrator` reads `ARES_BLUE_SIMULATED_CONTAINMENT=1` at startup; when set, it connects a NATS-backed recorder and logs confirmation; when unset or when NATS is unavailable, it falls back to a disabled recorder with a warning — blue detection and tracing are unaffected either way - `has_pending_cross_forest_escalation` now accepts `dominated_domains` and skips escalation vulns whose `target_domain` forest root is already dominated, closing the runaway-op bug without requiring a `written_off` stamp - `Dispatcher` gains a `red_draining: Arc<AtomicBool>` flag; `mark_red_draining` is called by the completion monitor the instant the op is deemed complete, and `do_submit_outcome` checks it as the single choke point for all red task submission **Removed:** - Static `confirm_escalation` / `downgrade_escalation` handling from `dispatch_callback` — these callbacks now route through the new async handlers instead of the inline `Some(CallbackResult::TaskComplete { .. })` stubs

**Key Changes:** - Moved remote build directory from tmpfs (`/tmp`) to disk-backed `/var/tmp` to prevent incremental build failures caused by systemd cleanup sweeping cargo build artifacts - Fixed PostgreSQL install check to detect the server rather than the client, preventing silent failures when `postgresql-client` is already present without a running cluster **Changed:** - Remote build directory path - Changed `REMOTE_BUILD_DIR` from `/tmp/ares-build` to `/var/tmp/ares-build` in `.taskfiles/ec2/Taskfile.yaml`; `/tmp` on kali-ares is a 7.7G tmpfs subject to 10-day systemd-tmpfiles-clean sweeps that were reaping cargo build-script outputs (e.g. `rustversion`'s `OUT_DIR/version.expr`) while leaving fingerprints intact, causing `include!(OUT_DIR/version.expr)` to fail with ENOENT on the next incremental build; `/var/tmp` is disk-backed with 53G free and a 30-day retention policy - PostgreSQL install trigger - Replaced `command -v psql` with `command -v pg_lsclusters` in `setup-history-db.sh`; kali-ares ships `postgresql-client` as a transitive dependency of other tooling, making `psql` available even when no server or cluster exists, so the old guard silently skipped installation; `pg_lsclusters` (from `postgresql-common`) is the exact binary the script depends on next and reliably indicates whether the server package is present

…260) **Key Changes:** - Report generation now runs on the remote backend (EC2 or K8s) rather than locally, fixing a fundamental issue where Blue state lives in the orchestrator's remote Redis, not localhost - Generated reports are automatically fetched back to the local machine via SSM (EC2) or `kubectl exec` (K8s) after remote generation - Added `REGENERATE` flag support to allow forcing report regeneration via `--regenerate` CLI argument **Added:** - `REGENERATE` variable support - new optional `REGENERATE=true` task parameter that passes `--regenerate` to the CLI, enabling forced report regeneration without changing the operation ID - Remote-to-local report fetch logic - after remote generation, the report path is parsed from CLI output and the file is pulled back locally using either SSM (`run_ssm_cmd`) for EC2 or `kubectl exec` for K8s transports, with validation that the fetched file is non-empty **Changed:** - Report generation transport - replaced direct local `{{.ARES_CLI}} blue report` invocation with `{{.ARES_CLI}} {{.TRANSPORT_ARGS}} blue report`, ensuring the command runs against the correct remote backend - Output directory structure - reports are now saved under `{{.OUTPUT_DIR}}/blue/` (with `mkdir -p` inlined into the main shell block) instead of directly under `{{.OUTPUT_DIR}}` - Shell robustness - added `set -euo pipefail` to the main command block to catch errors early and prevent silent failures - Task description - updated to document the new `REGENERATE=true` usage option

…ons (#261) **Key Changes:** - Introduces a pre-LLM deterministic sweep that runs the full detection catalog before the orchestrator loop starts, ensuring catalog coverage is no longer dependent on the LLM's token/context budget - Fixes multi-technique correlation by expanding `load_investigation_report` to return one `BlueTeamDetection` per distinct technique instead of collapsing to a single first match - Improves detection accuracy across several templates (golden ticket, S4U delegation, ADCS exploitation) and adds a new `detect_valid_account_reuse` template - Strengthens threat hunter guidance with mandatory attribution decision trees for S4U and valid-account reuse, and corrects the MITRE technique for constrained delegation from T1558.003 to T1550.003 **Added:** - Deterministic baseline sweep module (`sweep.rs`) - runs every detection template concurrently with bounded parallelism (default 6) under a configurable wall-clock cap (default 360s); records each hit as a MITRE technique, TTP-level evidence, and timeline event directly into blue state before the LLM loop begins; result is folded into the orchestrator task prompt so the LLM starts from a recorded baseline rather than rediscovering detections - `detect_valid_account_reuse` detection template targeting Event 4648 (explicit-credential logon) mapped to T1078.002, with patterns for `TargetServerName`, `TargetInfo`, `explicit.credential`, and `runas`, excluding machine-account targets - `SweepOutcome` and `FiredDetection` structs with a `prompt_summary()` method that generates a directive prompt section listing fired techniques, no-match templates, and any templates cut off by the time cap - Environment variable controls: `ARES_BLUE_DETERMINISTIC_SWEEP` (toggle, default on), `ARES_BLUE_SWEEP_CONCURRENCY`, and `ARES_BLUE_SWEEP_TIMEOUT_SECS` - Unit tests covering `parse_fire_count`, confidence/severity mapping, evidence type mapping, `sweep_enabled` toggle behavior, and `prompt_summary` output for both clean and timed-out runs **Changed:** - `load_investigation_report` return type changed from `Option<BlueTeamDetection>` to `Vec<BlueTeamDetection>`, scanning only the blue-authored body (before `## Appendix`) to avoid falsely crediting blue with red's ground-truth technique list embedded in the appendix; call site updated to use `extend` instead of conditional push - False-positive identification logic tightened to only flag a detection as a false positive when no red activity matches its technique, preventing parent/sub-technique pairs (e.g. T1021 and T1021.002) from spuriously flagging each other - `detect_s4u_delegation` filter corrected to a single `TransmittedServices` stage - the second stage filtering on service-class prefixes (`cifs/`, `ldap/`, etc.) was matching against the SAM account name in `ServiceName` rather than an SPN string, dropping ~98% of true positives (163 populated-TransmittedServices events reduced to 2 in live measurement); machine-account exclude removed for the same reason - `detect_golden_ticket` reworked to target Event 4769 only with RC4 encryption (`0x17`) on DC SPNs (`cifs/`, `ldap/`, `host/`, `krbtgt`), replacing tool-string patterns (`golden.*ticket`, `ticketer`) that never appear in Windows Security logs - `detect_adcs_exploitation` filter stage 2 extended with dangerous-EKU OIDs (`2.5.29.37.0` Any Purpose, `1.3.6.1.4.1.311.20.2.1` Certificate Request Agent, `1.3.6.1.4.1.311.76.6.1` SubCA) to catch ESC2/ESC3 enrollments that lack ESC keyword or SAN strings - Threat hunter prompt updated with mandatory attribution decision trees for S4U (T1550.003) and valid-account reuse (T1078.002), explicit instruction to run `detect_s4u_delegation` and `detect_valid_account_reuse` first before the token budget runs low, and a reinforced rule that a fired detection query must be recorded as its MITRE technique rather than described in prose

…ed filters (#262) **Key Changes:** - Fixed critical false positive flood in kerberoasting (690/870 events, 4% precision) and AS-REP roasting (411/590 events, 70% FP rate) by anchoring patterns to specific XML field names instead of spanning across field names and values - Replaced chained `|=` (conjunctive AND) with regex alternation `|~ "(?i)(…)"` for multi-literal pattern stages, fixing a logic bug that caused OR-intended stages to require all terms simultaneously - Introduced `TicketEncryptionType..u003e0x17` and `PreAuthType..u003e0.u003c` as field-anchored filter patterns, verified against 24h live Loki windows - Added regression tests covering the OR-vs-AND compile behavior and all three fixed detection rules **Added:** - Regression test `multi_literal_stages_or_not_and` - verifies that RBCD delegation attribute casings and remote-registry service state alternatives compile to a single regex OR rather than a conjunctive `|=` chain, guarding against re-introduction of the AND bug - Regression test `golden_ticket_keys_on_ticket_encryption_type` - asserts golden ticket stage 1 contains `TicketEncryptionType` and does not match bare `0x17`/`rc4` - Regression test `kerberoasting_keys_on_ticket_encryption_type` - asserts kerberoast uses the field-anchored pattern and excludes the name-spanning `encryption.*type` and SPN-shaped `servicename` patterns - Regression test `asrep_roasting_keys_on_preauthtype_zero` - asserts AS-REP roasting matches the closing-tag-anchored `PreAuthType..u003e0.u003c` and excludes all three previously over-broad patterns **Changed:** - Kerberoasting detection (`detect_kerberoasting`) - replaced three broken regex patterns (`encryption.*type.*(0x17|rc4)`, `ticket.*encryption.*(0x17|rc4)`, `servicename.*(mssql|http|ldap|cifs)`) with a single `TicketEncryptionType..u003e0x17` field-anchored filter; live data showed the old patterns matched 690/870 events with only 28 real RC4 tickets (662 false positives/day) - AS-REP roasting detection (`detect_asrep_roasting`) - replaced four over-broad OR'd patterns with `PreAuthType..u003e0.u003c`, anchoring the closing XML tag to prevent `preauthtype.*0` from matching PreAuthType values of 2/15/16/17 via trailing zeroes elsewhere on the line; live data showed 411/590 events fired where only 12 were genuine no-pre-auth TGTs - Golden ticket detection (`detect_golden_ticket`) - updated stage 1 from bare `['0x17', 'rc4']` to `['TicketEncryptionType..u003e0x17']`, preventing `rc4` from matching capability-enumeration fields (ServiceSupportedEncryptionTypes, ClientAdvertizedEncryptionTypes) present on ~90% of all 4769 events; stage 2 DC service class filter retained with expanded comment explaining why it currently zeroes recall and why it cannot be dropped without conflating golden ticket with kerberoasting - `build_pattern_filter` logic - removed the 2-3 literal fast-path that chained `|= "a" |= "b"` (which ANDs terms, requiring all to appear on one line); multi-literal stages now always compile to `|~ "(?i)(a|b)"` so patterns within a stage are correctly disjunctive; single literals retain the fast `|=` contains path - Pattern filter test (`pattern_filter_uses_contains_for_few_literals`) - renamed to `pattern_filter_ors_multiple_literals` and updated assertion from chained `|=` to regex alternation to match corrected behavior

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | returntocorp/semgrep | container | digest | `2b33f46` → `98c2572` | --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Enabled. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [docker/login-action](https://redirect.github.com/docker/login-action) ([changelog](https://redirect.github.com/docker/login-action/compare/af1e73f918a031802d376d3c8bbc3fe56130a9b0..abd2ef45e78c5afb21d64d4ca52ee8550d9572c7)) | action | digest | `af1e73f` → `abd2ef4` | --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Enabled. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [taiki-e/install-action](https://redirect.github.com/taiki-e/install-action) ([changelog](https://redirect.github.com/taiki-e/install-action/compare/a6b2e2dcd845ddd7f509ce4f3ed3d922b80cc5d9..3d7d7cd5ac7f994c1892ae0c06165095b9139094)) | action | digest | `a6b2e2d` → `3d7d7cd` | --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Enabled. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

| datasource | package | from | to | | ----------- | -------------------- | ------- | ------- | | github-tags | github/codeql-action | v4.37.2 | v4.37.3 |

This PR contains the following updates: | Package | Update | Change | |---|---|---| | grafana/loki | patch | `3.7.3` → `3.7.4` | --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Update | Change | |---|---|---| | [grafana/mimir](https://redirect.github.com/grafana/mimir) ([source](https://redirect.github.com/grafana/mimir/tree/HEAD/cmd/mimir)) | patch | `3.1.3` → `3.1.4` | --- ### Release Notes <details> <summary>grafana/mimir (grafana/mimir)</summary> ### [`v3.1.4`](https://redirect.github.com/grafana/mimir/blob/HEAD/CHANGELOG.md#314) ##### Grafana Mimir - \[BUGFIX] Packaging: Fix the DEB/RPM packages shipping the `mimir`, `mimirtool`, `metaconvert`, and `query-tee` binaries without the executable bit set, which caused `mimir.service` to fail to start. [#16166](https://redirect.github.com/grafana/mimir/issues/16166) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [tera](https://redirect.github.com/Keats/tera) | workspace.dependencies | minor | `2.0.0` → `2.1.0` | --- ### Release Notes <details> <summary>Keats/tera (tera)</summary> ### [`v2.1.0`](https://redirect.github.com/Keats/tera/blob/HEAD/CHANGELOG.md#210-2026-06-23) [Compare Source](https://redirect.github.com/Keats/tera/compare/v2.0.0...v2.1.0) - Add .iter() method to Kwargs - Add `Tera::contains_component` and `Tera::get_component_names` </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [base64](https://redirect.github.com/marshallpierce/rust-base64) | dependencies | minor | `0.22` → `0.23` | --- ### Release Notes <details> <summary>marshallpierce/rust-base64 (base64)</summary> ### [`v0.23.0`](https://redirect.github.com/marshallpierce/rust-base64/blob/HEAD/RELEASE-NOTES.md#0230) [Compare Source](https://redirect.github.com/marshallpierce/rust-base64/compare/v0.22.1...v0.23.0) - Added more consts for preconfigured configs and engines - Make DecodeError::InvalidLastSymbol more clear by including the decoded value - Added SIMD-accelerated engines behind the default-on `simd-unsafe` feature: `Simd` picks the best instruction set at runtime (AVX2 on `x86_64`, NEON on `aarch64`) and falls back to the scalar `GeneralPurpose` engine, while `Avx2` and `Neon` target one instruction set with no runtime detection and work in `no_std`. The engines support the standard and URL-safe alphabets. - Update MSRV to 1.71.0 - Add support for custom padding symbols </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Update | Change | |---|---|---| | [grafana/grafana](https://redirect.github.com/grafana/grafana) | patch | `13.1.2` → `13.1.3` | --- ### Release Notes <details> <summary>grafana/grafana (grafana/grafana)</summary> ### [`v13.1.3`](https://redirect.github.com/grafana/grafana/releases/tag/v13.1.3): 13.1.3 [Download page](https://grafana.com/grafana/download/13.1.3) [What's new highlights](https://grafana.com/docs/grafana/latest/whatsnew/) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate CLI](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Update | Change | |---|---|---| | grafana/loki | patch | `3.7.4` → `3.7.6` | --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate CLI](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [async-trait](https://redirect.github.com/dtolnay/async-trait) | dev-dependencies | patch | `0.1.91` → `0.1.92` | | [async-trait](https://redirect.github.com/dtolnay/async-trait) | dependencies | patch | `0.1.91` → `0.1.92` | --- ### Release Notes <details> <summary>dtolnay/async-trait (async-trait)</summary> ### [`v0.1.92`](https://redirect.github.com/dtolnay/async-trait/releases/tag/0.1.92) [Compare Source](https://redirect.github.com/dtolnay/async-trait/compare/0.1.91...0.1.92) - Resolve double\_must\_use clippy lint in generated code ([#303](https://redirect.github.com/dtolnay/async-trait/issues/303)) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about these updates again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate CLI](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

This PR contains the following updates: | Package | Type | Update | Change | |---|---|---|---| | [thiserror](https://redirect.github.com/dtolnay/thiserror) | workspace.dependencies | patch | `2.0.19` → `2.0.20` | --- ### Release Notes <details> <summary>dtolnay/thiserror (thiserror)</summary> ### [`v2.0.20`](https://redirect.github.com/dtolnay/thiserror/releases/tag/2.0.20) [Compare Source](https://redirect.github.com/dtolnay/thiserror/compare/2.0.19...2.0.20) - Suppress redundant\_field\_names clippy lint in generated code ([#454](https://redirect.github.com/dtolnay/thiserror/issues/454)) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate CLI](https://redirect.github.com/renovatebot/renovate).  Co-authored-by: ares-renovate[bot] <286782180+ares-renovate[bot]@users.noreply.github.com>

**Key Changes:** - Stripped hundreds of obvious, self-explanatory inline comments throughout the ares-cli, ares-core, ares-llm, and ares-tools crates to reduce noise and improve readability - Rephrased a few remaining comments to be more concise and meaningful rather than restating the code - No functional or behavioral changes — this is a pure cleanup pass **Changed:** - Removed redundant "step-describing" comments across orchestrator automation modules, state persistence, result processing, and dispatcher logic that merely narrated what the following line already made clear - Cleaned up section-label comments in report generators (blue team, red team, detection playbooks, correlation reports) that duplicated the associated `println!`/`push_str` section headers - Trimmed obvious comments from parsers (nmap, secretsdump, shares, trust, ntsd, credential tools) and blue team investigation read/write helpers where the code was already descriptive - Condensed retained comments for clarity — e.g., reworded the exploitation "permanently marked exploited" note to "Set by result processing on success" and simplified the MOTD banner-frame comment in `ares-tools/src/filter.rs` - Removed narrative comments from the LLM agent loop, providers (anthropic, openai), routing, and tool registry, and folded the "persists to Redis" note directly onto the relevant token-usage callback

…sing (#482) **Key Changes:** - Routed MSSQL NTLM coercion tasks to the `privesc` role since only that role registers the `mssql_ntlm_coerce` tool - Made the coercion prompt handle singular `technique` keys and `relay_target` fallbacks so more task payloads render correctly - Hardened `mssql_ntlm_coerce` output parsing to only claim a coercion attempt when `xp_dirtree` actually executed, eliminating false positives from listener-side captures and connection failures **Added:** - `impacket-mssqlclient` to the expected worker tool set so MSSQL coercion has its required binary - `ares-cli/src/worker/tool_check.rs` - `xp_dirtree_executed` helper that detects real execution by checking for the `subdirectory` result marker while rejecting known failure markers (login failed, permission denied, connection refused, TDS connect, timeout) - `ares-tools/src/parsers/mod.rs` - Extensive test coverage for coercion routing, prompt rendering variants (singular vs plural techniques, relay-target fallback, unknown targets), and parser behavior across empty output, refused logins, denied procedures, TDS failures, and listener-side vs result-set captures **Changed:** - MSSQL coercion dispatch now targets a named `MSSQL_COERCION_TARGET_ROLE` (`privesc`) constant instead of the hardcoded `"coercion"` role - `ares-cli/src/orchestrator/automation/mssql_coercion.rs` - Coercion prompt generation now resolves the target via `coercion_target` (with `relay_target` fallback) and techniques via `coercion_techniques` (accepting a singular `technique` key when the plural array is absent), with plural taking precedence - `ares-llm/src/prompt/coercion.rs` - `mssql_ntlm_coerce` parsing now emits the `coercion_attempted` vulnerability marker only when `xp_dirtree_executed` confirms execution, gating it behind real result-set evidence - `ares-tools/src/parsers/mod.rs` **Removed:** - Listener-side NetNTLMv2 hash extraction from the `mssql_ntlm_coerce` arm, since raw hash captures are handled by listener-side parsers rather than the coercion tool's output - `ares-tools/src/parsers/mod.rs`

**Key Changes:** - Added underscore-formatted IP detection to the GOAD token sweep script to catch lab IPs that use underscores instead of dots - Updated a display test to use a valid lab IP pattern matching the underscore convention **Added:** - Underscore IP pattern matching - Introduced an `ips_underscore` pattern and appended it to the `banned` regex in `scripts/goad-token-sweep.sh` so IPs like `10_1_2_51` are detected alongside their dotted equivalents **Changed:** - Token category test assertion - Updated the `token_category` test in `ares-cli/src/ops/loot/format/display.rs` to assert on `mssql_192_168_58_51` instead of `mssql_10_1_2_51`, aligning the test fixture with the underscore IP format

…ckends (#484) **Key Changes:** - Added `task ec2:e2e`, a gated end-to-end runner that deploys, verifies binary provenance, runs a full red(+blue) operation, and fetches both reports - Unified all `blue:*` tasks behind a single `BLUE_TRANSPORT` selector (`ec2`, `k8s`, or `local`) and consolidated the alert-submission tasks into one `blue:submit` - Removed the K8s pod file-sync tasks (`remote:sync`, `remote:sync:full`) and red-team replay recording tasks in favor of binary-deploy workflows - Reworked the Rust cross-compile path to prefer `cargo-zigbuild` on all hosts, avoiding rustc crashes under qemu emulation on Apple Silicon **Added:** - End-to-end operation runner - Introduced `task ec2:e2e` and its backing `.taskfiles/ec2/scripts/e2e-op.sh`, which sanity-checks the repo, gates against stale source and non-matching binary SHAs, refuses prod hosts, restarts workers to drop the poisoned tool cache, launches a fresh op, scans for the tool-pruning-cascade regression, and fetches red/blue reports - Log filtering on EC2 - Added a `FILTER` var to `ec2:logs` with `grep --line-buffered` so blue lines can be isolated from the interleaved orchestrator log without buffering delays - Proxmox taskfile include - Wired `.taskfiles/proxmox/Taskfile.yaml` into the root `Taskfile.yaml` - Documentation for the new workflows - Expanded `README.md`, `docs/blue.md`, and `docs/benchmark-replay.md` to cover `ec2:e2e`, `blue:submit`, transport-aware logs, and the corrected `CAPTURE`/build guidance **Changed:** - Blue transport model - Replaced the ec2/k8s if-branch in `.taskfiles/blue/Taskfile.yaml` with a `case` supporting a new `local` transport, and made `blue:submit`, `blue:playbook`, and `blue:multi:logs` transport-aware; the playbook now streams JSON over the transport instead of relying on a `kubectl cp` that silently copied nothing - Secure alert submission - `blue:submit` now passes the alert by value and drops `--grafana-api-key`, falling back to the remote's own `GRAFANA_SERVICE_ACCOUNT_TOKEN` to avoid leaking it into SSM history and CloudTrail - Rust cross-compile ordering - Reordered `remote:rust:build` in `.taskfiles/remote/Taskfile.yaml` to try `cargo-zigbuild` first and warn when falling back to `cross` on arm64 macOS - Grafana token field lookup - Updated the `GRAFANA_SERVICE_ACCOUNT_TOKEN` 1Password field from `api-token` to `grafana-token` - Pre-commit task reference - Repointed the CI workflow and root taskfile to the namespaced `pre-commit:*` tasks (`pre-commit:run-pre-commit`, `pre-commit:install-pc-hooks`) - MAX_STEPS is now overridable - `blue:submit` and `blue:multi:remote` accept a `MAX_STEPS` override instead of a hardcoded value - README key-tasks table - Refreshed the blue task table and added a note that every `blue:*` task reads `BLUE_TRANSPORT` **Removed:** - K8s file-sync tasks - Deleted `remote:sync` and `remote:sync:full` and their supporting vars (`PARALLELISM`, `PVC_PATH`, `VERIFY_PVC_DIFF`) from `.taskfiles/remote/Taskfile.yaml`, shifting to binary deploys - Red-team replay recording tasks - Removed `red:multi:replay:copy`, `:cat`, `:list`, and `:clear` from `.taskfiles/red/Taskfile.yaml` - Redundant blue tasks - Dropped `blue:once:remote`, `blue:investigate`, and `blue:multi` in favor of the unified `blue:submit` - Obsolete root taskfile targets - Removed `get-dotenv-value`, `run-pre-commit`, `rust:clean`, `ares:config:show`, and `check-aws-auth` from `Taskfile.yaml`

…tput (#485) **Key Changes:** - Added an early return in `parse_tool_output` to short-circuit when tool output is empty, preventing parser arms from fabricating evidence purely from input parameters - Introduced a regression test that verifies no parser arm mints evidence keys from params alone across all 78 supported tools - Added a guard test to keep the evidence key list synchronized with the orchestrator's `result_has_parser_evidence` gate **Added:** - Empty-output guard in `parse_tool_output` - Returns an empty discoveries object immediately when the trimmed output is empty, closing a path where the orchestrator could accept params-derived markers as proof an exploit succeeded (`ares-tools/src/parsers/mod.rs`) - Comprehensive false-evidence regression test - Added `no_parser_arm_mints_evidence_from_params_alone`, which iterates every entry in the new `PARSED_TOOLS` list with a fully populated params object and empty output, asserting no `EVIDENCE_KEYS` array is populated; offenders are reported with context explaining the security impact (`ares-tools/src/parsers/mod.rs`) - Orchestrator gate consistency test - Added `evidence_key_list_matches_the_orchestrator_gate` to assert the `EVIDENCE_KEYS` count stays aligned with `result_has_parser_evidence` in `ares-cli/src/orchestrator/result_processing/mod.rs`, prompting both to be updated together (`ares-tools/src/parsers/mod.rs`) - Test fixtures and constants - Added the `EVIDENCE_KEYS` and `PARSED_TOOLS` slices plus the `fully_populated_params` helper to support the new tests (`ares-tools/src/parsers/mod.rs`)

…sk consolidation (#486) **Key Changes:** - Consolidated blue-team tasks: replaced `blue:investigate`/`blue:multi` with `blue:submit`, absorbed `blue:once:remote` into `blue:multi:remote`, and made `blue:multi:logs`/`blue:playbook` transport-aware - Removed the dead `red:multi:replay:*` task surface and its documentation, and the Python-era `remote:sync`/`remote:sync:full` tasks **Changed:** - Blue task surface documentation in `blue-team.md` reflects `blue:submit` now honoring `{{.TRANSPORT_ARGS}}` (agreeing with read tasks), `blue:multi:logs` becoming transport-aware (tailing `/var/log/ares/orchestrator.log` on EC2), `blue:playbook` rewritten to stream JSON over stdout, and `BLUE_TRANSPORT` accepting `local` - Replay documentation in `benchmarks-and-replay.md` rewritten to record that the four k8s-only `red:multi:replay:*` tasks were removed since nothing ever wrote `/ares/replay/recording.jsonl` - Config and CI references updated: `ares:config:show` folded into `ares:config:check` (`config-and-env.md`), CI repointed from the duplicate root `run-pre-commit` to `pre-commit:run-pre-commit`, and `task init` fixed to call `pre-commit:install-pc-hooks` (`tools-and-gates.md`) - Detection playbook output path moved to `blue/<op_id>_detection_playbook.json` and the `e2e-op.sh` knobs list expanded with `ALLOW_STALE`, `BLUE_SETTLE_WAIT`/`BLUE_STALL_WAIT`, and stricter `BLUE=0`/`SKIP_RESTART` handling - Gemini operator agent examples updated to use `blue:multi:remote` and the new `blue:submit ALERT=...` task (`.gemini/agents/ares-operator.md`) **Removed:** - The `red:multi:replay:*` task family (`copy`, `cat`, `list`, `clear`) and its destructive-task warning, plus the `recordings/` output directory notes describing the always-empty dead surface - The Python-era `remote:sync` and `remote:sync:full` tasks that operated on the long-gone `src/ares/**` tree, with guidance now pointing at `task -y remote:rust:build && task -y remote:rust:deploy TEAM=blue` - Stale `BUILD_TOOL` defaults-to-`auto` claims and the duplicate root-level `run-pre-commit` CI task

#488) **Key Changes:** - Replaced hardcoded S3 buckets, regions, and account IDs with runtime-injected placeholders and generic references across build scripts and docs - Renamed the ambiguous "phase2" golden AMI build stage to the descriptive "tools-install" oneshot service throughout - Made RDS endpoint configuration explicit and env-driven rather than defaulting to a hardcoded RDS host **Changed:** - Golden AMI userdata parameterization - Replaced hardcoded `BUCKET` and `us-east-1` region values in `scripts/ares-golden-userdata.sh` with `__BUCKET__` and `__AWS_REGION__` placeholders that are rendered at build time, so the script is no longer tied to a specific account or region - Build-time placeholder rendering - Updated `scripts/build-ares-golden-ami.sh` to `sed`-substitute the placeholders into a temp userdata file (with a cleanup trap and a guard that fails if any placeholder is left unrendered), and pass that rendered file to `run-instances` - Phase renaming - Renamed the post-reboot `ares-phase2.service`/`ares-phase2.sh` to `ares-golden-tools-install.service`/`ares-golden-tools-install.sh`, and the `PHASE2_DONE` S3 completion marker to `TOOLS_INSTALL_DONE`, updating all references, log messages, and comments for clarity in both scripts - RDS endpoint handling - Changed `.taskfiles/ec2/Taskfile.yaml` so `RDS_ENDPOINT` now defaults from the `ARES_RDS_ENDPOINT` env var (or empty) instead of a hardcoded RDS host, and gated the Secrets Manager password lookup on `RDS_ENDPOINT` being supplied - Documentation genericization - Replaced the literal AWS account ID `381491903301` with `<account-id>` in the staging bucket examples in `docs/benchmark-replay.md` and `warpgate-templates/templates/ares-replay-stack/README.md`

**Key Changes:** - Extracted the inline attacker IP resolution logic from the Proxmox Taskfile into a dedicated, reusable shell script - Replaced the `sh:`-computed `ATTACKER_IP` variable with an `ATTACKER_IP_CMD` invocation, deferring IP resolution to runtime in each task - Updated all tasks to resolve the attacker IP on demand via command substitution rather than relying on a pre-computed value **Added:** - Standalone attacker IP resolution script - Created `.taskfiles/proxmox/scripts/attacker-ip.sh`, which queries the Proxmox guest agent for network interfaces and parses out the first non-loopback IPv4 address via Python, accepting `PROXMOX_SSH_HOST` and `ATTACKER_VMID` as positional arguments with environment-variable fallbacks **Changed:** - IP resolution strategy in `.taskfiles/proxmox/Taskfile.yaml` - Replaced the eagerly evaluated `ATTACKER_IP` variable (which ran the guest-agent query at parse time) with an `ATTACKER_IP_CMD` pointer to the new script, so the IP is resolved lazily per task via `IP="$({{.ATTACKER_IP_CMD}})"`; this avoids running the query on every task invocation and keeps the complex parsing logic out of the YAML **Removed:** - Inline IP resolution logic - Removed the embedded multi-line Python guest-agent parsing block from the `ATTACKER_IP` var definition in `.taskfiles/proxmox/Taskfile.yaml`, now superseded by the external script

**Key Changes:** - Removed all compiled-in S3 bucket, region, and profile defaults from the benchmark tooling, since these values are account-specific and previously leaked lab/dev account names into the binary - Introduced a new `benchmark:` configuration section and shared resolution helpers so settings can come from `BENCHMARK_*` / `LOKI_S3_*` env vars or `ares.yaml`, failing loudly when required values are missing - Made bucket and region required inputs with clear error messages naming both the env var and config key, rather than silently falling back to a hardcoded account **Added:** - New `BenchmarkConfig` section - Added an optional `benchmark:` config struct (`s3_bucket`, `aws_profile`, `aws_region`, `loki_s3_bucket`, `loki_s3_region`, `loki_s3_profile`) in `ares-core/src/config/sections.rs`, wired into `AresConfig` in `config/mod.rs` - Shared resolution helpers - Added `benchmark_section`, `resolve`, and `require` in `snapshot_s3.rs`; `resolve` treats empty as unset (defaulting to the credential chain) while `require` errors when a value with no sane fallback is missing - New env vars and validation - Added `LOKI_S3_BUCKET`, `LOKI_S3_REGION`, `LOKI_S3_PROFILE` and a `BENCHMARK_S3_BUCKET` precondition check in `.taskfiles/benchmark/Taskfile.yaml` **Changed:** - Config resolution across benchmark commands - Updated `LokiS3::from_env`, `SnapshotConfig::from_env`, and their callers in `capture.rs`, `mod.rs`, and `replay.rs` to return `Result` and pull values from env or config instead of hardcoded constants - Documentation and examples - Updated `.env.example`, `docs/benchmark-replay.md`, and `README.md` to mark `BENCHMARK_S3_BUCKET`, `BENCHMARK_AWS_REGION`, `LOKI_S3_BUCKET`, and `LOKI_S3_REGION` as required with no defaults, and annotated the omitted `benchmark:` section in `config/ares.yaml` - Taskfile default bucket - Changed the `S3_BUCKET` var in `.taskfiles/benchmark/Taskfile.yaml` to have no default, requiring it via `.env` - Golden userdata ordering - Reordered `scripts/ares-golden-userdata.sh` to disable the tools-install service before writing the completion marker to S3 **Removed:** - Hardcoded S3 constants - Removed `DEFAULT_LOKI_S3_BUCKET/REGION/PROFILE` and `DEFAULT_BENCHMARK_BUCKET/PROFILE/REGION` from `capture.rs`, plus `DEFAULT_S3_BUCKET`, `DEFAULT_AWS_REGION`, and `DEFAULT_AWS_PROFILE` from `snapshot_s3.rs`, eliminating account-specific values baked into the binary

**Key Changes:** - Introduced a shared `CrackInflight` guard on `Dispatcher` so the automation tick and the LLM `dispatch_crack` tool can no longer independently re-queue the same hash, which previously wedged both hashcat GPU slots for ~50 minutes across five redundant runs - Replaced the per-loop local dedup `HashMap` and atomic counter in `auto_crack_dispatch` with atomic capacity-checked reservations that a second racing producer cannot bypass - Added LLM-facing refusal/deferral responses in the `dispatch_crack` tool for hashes that are already in flight, at capacity, or have exhausted their crack attempts - Wired crack-slot release into the task-completion and stale-task reaping paths so a dying dispatch self-heals via a TTL backstop instead of blocking a hash permanently **Added:** - Shared crack-scheduling guard - New `CrackInflight` type in `dispatcher/mod.rs` with per-dispatch grouped reservations, a derived active-count, `CRACK_INFLIGHT_TTL` backstop (45m), and `DEFAULT_MAX_ACTIVE_CRACK_TASKS` cap configurable via `ARES_MAX_ACTIVE_CRACK_TASKS`; exposes `try_reserve`, `release`, `is_inflight`, `at_capacity`, `active`, `live_keys`, and `expire_stale` - LLM crack-tool guardrails - In `callback_handler/dispatch.rs`, added early `CallbackResult::Continue` responses that refuse hashes exceeding `MAX_CRACK_ATTEMPTS` or already given up, defer when all slots are busy, and reject re-dispatch of in-flight hashes, reserving the slot on successful dispatch - Reservation test coverage - Added async tests in `dispatcher/mod.rs` covering cross-producer blocking, cap enforcement, partial-batch reservation, idempotent release, and the TTL-vs-stall-ceiling invariant **Changed:** - Crack dispatch flow - `auto_crack_dispatch` now reserves candidate keys atomically before batching, releases the slot on both stall-timeout and completion, and derives active/inflight state from the shared guard rather than local structures - Task cleanup and completion paths - `release_reaped_task` and `cleanup_stale_tasks` in `monitoring.rs` and `process_completed_task` in `result_processing/mod.rs` now release the crack reservation for reaped or completed tasks - Heartbeat monitor signature - `spawn_heartbeat_monitor` and its call sites in `orchestrator/mod.rs` now take the full `Arc<Dispatcher>` instead of just `credential_inflight`, giving the monitor access to `crack_inflight` - Module exports - `automation/mod.rs` now re-exports `MAX_CRACK_ATTEMPTS` for use by the LLM crack tool **Removed:** - Local inflight tracking in `auto_crack_dispatch` - Removed the `InflightCrackSlot` drop-guard, the `inflight_crack_dedup` `HashMap`, the `AtomicUsize` task counter, the `max_active_crack_tasks` env helper, and the now-obsolete drop-on-panic slot-release test, all superseded by the shared `CrackInflight` guard

**Key Changes:** - Prevent the planner from re-dispatching WinRM lateral movement against targets whose backing vulnerability has hit the max exploit-failure abandonment cap - Added per-target resolution of the WinRM `winrm_access` vulnerability that backs a lateral technique - Added comprehensive unit tests covering positive, negative, and cross-host cases **Added:** - WinRM backing vuln resolution - Introduced `lateral_backing_vuln_id` in `task_builders.rs` to locate the `winrm_access` vulnerability matching a given target IP, returning `None` for non-WinRM techniques so unrelated techniques are never gated - Abandonment guard in the dispatcher - Added a check that skips lateral dispatch (returning `Ok(None)` with a debug log) when the backing vuln has been abandoned at max exploit failures, stopping the planner from repeatedly targeting a dead WinRM host - Test coverage - Added three tests in `task_builders.rs` verifying the backing vuln resolves for a matching WinRM target, is absent for non-WinRM techniques like psexec, and is scoped per target so one dead host does not suppress others

…ue hole (#492) **Key Changes:** - Moved crack dedup key reservation from the callback dispatch path to the submission layer, ensuring every crack task id — including deferred re-submits from the drain — is tracked - Added a non-refusing `reserve` method to `CrackInflight` so a real hashcat run can never be left untracked, closing the hole that allowed the same hash to be queued twice - Introduced `crack_dedup_key_from_payload` to derive dedup keys directly from the submitted payload rather than the originating `Hash` **Added:** - Payload-based dedup key derivation - Added `crack_dedup_key_from_payload` and a shared `crack_dedup_key_parts` helper in `automation/mod.rs`, letting the submission layer compute keys from a payload without access to the source `Hash` - Unconditional reservation recording - Added `CrackInflight::reserve` in `dispatcher/mod.rs`, which always records a reservation (unlike `try_reserve`), allowing `active` to briefly exceed `max_active` to reflect true state instead of a capped guess - Submission-layer reservation - Wired the new reservation into `submission.rs` so every crack task id is reserved as it is submitted - Test coverage - Added tests verifying payload-derived keys match those built from the original `Hash` and that non-crack payloads yield `None` **Changed:** - Dedup key computation - Refactored `crack_dedup_key` to delegate to `crack_dedup_key_parts`, and switched prefix slicing to be char-based (`chars().take(32)`) to avoid byte-boundary issues **Removed:** - Callback-path reservation - Removed the `try_reserve` call from `dispatch.rs`, since reservation now happens centrally at the submission layer where every task id is observed

**Key Changes:** - Introduced a `reasoning_effort` setting that threads from YAML/env config through the agent loop into OpenAI reasoning-model requests - Enforced strict precedence (env override > YAML > provider default) with validation that silently ignores unrecognised values to avoid 400 errors on every call - Restricted forwarding of `reasoning_effort` to gpt-5 family models only, since non-reasoning models reject the parameter outright **Added:** - Reasoning effort configuration field - Added `reasoning_effort: Option<String>` to `AgentConfig` in `ares-core/src/config/sections.rs` and defaulted it to `None` in config construction paths (`ares-core/src/config/mod.rs`) - Config layering and validation - Added `with_config_reasoning_effort` builder plus a `normalize_reasoning_effort` helper in `ares-llm/src/agent_loop/config.rs` that trims, lowercases, and accepts only `minimal`/`low`/`medium`/`high`, honouring the `ARES_AGENT_REASONING_EFFORT` env override; unrecognised values are dropped with a warning - Request plumbing - Added `reasoning_effort` to `AgentLoopConfig` and `LlmRequest` (`ares-llm/src/provider/mod.rs`), wired it through `run_agent_loop_inner` in `ares-llm/src/agent_loop/runner.rs` - OpenAI provider gating - Added `supports_reasoning_effort` and a `reasoning_effort` field on `ApiRequest` in `ares-llm/src/provider/openai.rs`, forwarding the value only to gpt-5 family models and omitting it from the wire when unset - Per-role defaults - Set `reasoning_effort: low` for the recon, credential_access, cracker, lateral_movement, and coercion agents in `config/ares.yaml` - Test coverage - Added precedence and casing tests for config layering and provider tests verifying gating and wire omission **Changed:** - Per-role model logging - Extended the orchestrator's per-role model log line in `ares-cli/src/orchestrator/mod.rs` to include the effective reasoning effort (falling back to `provider-default`)

…494) **Key Changes:** - Fixed crack tasks collapsing onto a single queue signature because their payloads carry no technique, target_ip, credential, or finding fields, causing distinct hashes to be silently dropped while `enqueue` reported them as successfully queued - Routed crack task identity through the same `crack_dedup_key_from_payload` used by the in-flight reservation guard, keeping queue-identity and run-identity aligned - Added comprehensive test coverage verifying signature behavior across distinct, duplicate, and non-crack task types **Added:** - Test suite for crack task signatures - Added `make_crack_task` helper and four tests in `ares-cli/src/orchestrator/deferred.rs` covering distinct-hash differentiation, genuine-duplicate collapsing (including case-insensitivity), reservation-key alignment, and confirmation that hash identity only affects crack tasks while other task types keep their existing signature behavior **Changed:** - Signature derivation for crack tasks - Modified `finding_key` in `ares-cli/src/orchestrator/deferred.rs` to derive the discriminating component from `crack_dedup_key_from_payload` when the task type is `crack`, prefixing the result with `crack|`, so two crack tasks that share a signature are exactly the two that would contend for one in-flight reservation - Documentation on `signature` and `finding_key` - Expanded doc comments to explain the crack-queue dedup gap (mirroring the prior ACL gap) and the reasoning behind unifying queue and run identity

**Key Changes:** - Removed `base_url` and `api_key` fields from Grafana configuration in favor of `dashboard_uid` - Ensured backward compatibility by gracefully ignoring legacy credential keys during config loading - Updated example configuration and documentation to reflect the streamlined Grafana section **Changed:** - Grafana config struct simplified - Removed `base_url` and `api_key` fields from `GrafanaConfig`, retaining only `enabled` and `dashboard_uid` in `ares-core/src/config/sections.rs` - Config loading resilience - Added `grafana_ignores_legacy_credential_keys` test to verify that configs containing the removed `base_url` and `api_key` keys still parse without error, and updated existing tests to assert `dashboard_uid` parsing in `ares-core/src/config/mod.rs` - Example configuration cleanup - Removed the `base_url` and `api_key` entries from the Grafana section in `config/ares.yaml` - Documentation update - Removed the outdated Grafana credential configuration snippet from the blue team observability docs in `docs/blue.md`

…ion gating (#496) **Key Changes:** - Fixed deployment label resolution so empty values fall through to the `ARES_DEPLOYMENT` env var instead of pinning LogQL queries to `deployment=""`, which matched nothing - Gated e2e operation consolidation on operation finalization (`completed_at`) rather than counts alone, preventing the terminal investigation from being dropped from the scorecard - Routed blue orchestrator tools in-process via a new `BlueToolDispatcher` **Added:** - Shared `resolve_deployment` helper in `investigation.rs` that treats empty alert labels and env vars as absent, correctly falling back from label to `ARES_DEPLOYMENT` — this fixes cases where an unset target environment yielded `Some("")` and suppressed the env fallback - Unit test `resolve_deployment_label_env_precedence` covering label/env precedence and empty-value fallthrough, deliberately consolidated into one test since parallel test execution races on process-global `std::env` - `BlueToolDispatcher` wiring in `run_investigation` to route orchestrator tools in-process, with an info log capturing the investigation ID and resolved deployment **Changed:** - Blue completion detection in `e2e-op.sh` now also reads `completed_at` from Redis and treats an unset value as blue still outstanding, since the terminal investigation is submitted minutes after `ops status` first reports "completed" - `BlueCallbackHandler::new` in `callbacks.rs` and `run_investigation` now delegate to the shared `resolve_deployment` helper instead of duplicating inline label/env extraction logic

…-main # Conflicts: # .github/workflows/build-and-push-templates.yaml # .github/workflows/meta-sync-labels.yaml # .github/workflows/molecule.yaml # .github/workflows/pre-commit.yaml # .github/workflows/release.yaml # .github/workflows/renovate.yaml # .github/workflows/rust.yaml # .github/workflows/semgrep.yaml # .github/workflows/test-template-builds.yaml # .github/workflows/validate-templates.yaml # Cargo.lock # Cargo.toml # ansible/requirements.yml # ares-cli/src/history/cost.rs # ares-cli/src/history/list.rs # ares-cli/src/history/search.rs # ares-cli/src/orchestrator/automation/adcs_exploitation.rs # ares-cli/src/orchestrator/automation/mssql_exploitation.rs # ares-cli/src/orchestrator/blue/chaining.rs # ares-cli/src/orchestrator/deferred.rs # ares-cli/src/orchestrator/diversity.rs # ares-cli/src/orchestrator/mod.rs # ares-cli/src/orchestrator/result_processing/mod.rs # ares-cli/src/orchestrator/strategy.rs # ares-core/src/config/defaults.rs # ares-core/src/config/sections.rs # ares-core/src/persistent_store/queries/credentials.rs # ares-llm/src/prompt/templates.rs # ares-llm/src/tool_registry/privesc/adcs.rs # ares-llm/templates/redteam/agents/system_instructions.md.tera # ares-tools/Cargo.toml # ares-tools/src/parsers/certipy.rs # ares-tools/src/parsers/delegation.rs # ares-tools/src/parsers/mssql.rs # config/ares.yaml # docs/attack-path-diversity.md

**Added:** - Molecule test workflow (`molecule.yaml`) with dynamic role change detection, matrix generation, and configurable parallelism for testing Ansible roles on PRs, pushes, and schedules **Changed:** - Renovate config (`renovate.json5`) to use branch-based automerge instead of PR-based, adding the `:automergeBranch` preset - Reverted action pins across all workflows to older versions - `actions/checkout` to v7.0.0, `actions/setup-python` to v6.3.0, `actions/setup-go` to v6, `dtolnay/rust-toolchain`, `taiki-e/install-action`, `softprops/action-gh-release` to v3.0.1, `actions/upload-artifact` to v7.0.0, and `renovatebot/github-action` to v46.1.20 - Reverted `WARPGATE_VERSION` from v4.10.0 to v4.9.1 across build, test, and validate workflows - Reverted Docker Buildx setup to use `docker/setup-buildx-action` directly with inline `driver-opts` instead of the composite action - Reverted `build-and-push-templates.yaml` triggers to drop Rust crate paths and scheduled cron, and simplified disk cleanup back to a minimal `rm -rf` - Reverted cache export to drop `ignore-error=true` from `--cache-to` across build workflows - Reverted `test-template-builds.yaml` change detection to templates-only, restored push-based build flow, and updated registry references to `ghcr.io/dreadnode/*` - Reverted `pre-commit.yaml` to drop remote Taskfile prefetch and Rust hook skipping, using `task -y run-pre-commit` - Reverted `semantic-prs.yaml` to drop `workflow_dispatch` manual validation path - Reverted `rust.yaml` clippy to run without `--all-targets` **Removed:** - Composite `setup-buildx` action (`.github/actions/setup-buildx/action.yml`) that retried BuildKit image pulls - `report-failure` job from `build-and-push-templates.yaml` that opened/updated failure issues on main - Proxmox placeholder env vars from `validate-templates.yaml` - `continue-on-error` from Semgrep SARIF upload step and `RENOVATE_FORK_PROCESSING` from the Renovate workflow

codecov · 2026-08-10T04:31:04Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.33%. Comparing base (dd71e14) to head (726c1e7).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #441      +/-   ##
==========================================
+ Coverage   79.97%   81.33%   +1.35%     
==========================================
  Files         434      476      +42     
  Lines      126175   170395   +44220     
==========================================
+ Hits       100913   138588   +37675     
- Misses      25262    31807    +6545

Files with missing lines	Coverage Δ
ares-cli/src/benchmark/capture.rs	`0.00% <ø> (ø)`
ares-cli/src/benchmark/manifest.rs	`100.00% <ø> (ø)`
ares-cli/src/benchmark/mod.rs	`0.00% <ø> (ø)`
ares-cli/src/benchmark/replay.rs	`9.55% <ø> (ø)`
ares-cli/src/benchmark/snapshot_s3.rs	`0.00% <ø> (ø)`
ares-cli/src/benchmark/tempo_push.rs	`66.37% <ø> (ø)`
ares-cli/src/blue/delete.rs	`0.00% <ø> (ø)`
ares-cli/src/blue/evidence.rs	`0.00% <ø> (ø)`
ares-cli/src/blue/mod.rs	`0.00% <ø> (ø)`
ares-cli/src/blue/operation.rs	`0.00% <ø> (ø)`
... and 42 more

... and 310 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

l50 and others added 30 commits July 20, 2026 22:05

chore(deps): update github/codeql-action action to v4.37.2 (#248)

bd999db

| datasource | package | from | to | | ----------- | -------------------- | ------- | ------- | | github-tags | github/codeql-action | v4.37.1 | v4.37.2 |

chore(deps): update actions/checkout action to v7.0.1 (#246)

a86354c

| datasource | package | from | to | | ----------- | ---------------- | ------ | ------ | | github-tags | actions/checkout | v7.0.0 | v7.0.1 |

chore(deps): update renovatebot/github-action action to v46.1.20 (#250)

482bedc

| datasource | package | from | to | | ----------- | ------------------------- | -------- | -------- | | github-tags | renovatebot/github-action | v46.1.19 | v46.1.20 |

chore(deps): update actions/labeler action to v7 (#254)

4d8de8f

| datasource | package | from | to | | ----------- | --------------- | ------ | ------ | | github-tags | actions/labeler | v6.2.0 | v7.0.0 |

chore(deps): update actions/setup-python action to v7 (#255)

4857567

| datasource | package | from | to | | ----------- | -------------------- | ------ | ------ | | github-tags | actions/setup-python | v6.3.0 | v7.0.0 |

chore(deps): update github/codeql-action action to v4.37.3 (#267)

679a33e

| datasource | package | from | to | | ----------- | -------------------- | ------- | ------- | | github-tags | github/codeql-action | v4.37.2 | v4.37.3 |

ares-renovate Bot and others added 22 commits August 8, 2026 19:19

l50 mentioned this pull request Aug 10, 2026

feat: sync upstream main with orchestrator, blue-team, and benchmark overhaul #439

Closed

dreadnode-renovate-bot Bot added area/docs Changes made to project documentation area/pre-commit Changes made to pre-commit hooks area/templates Changes made to warpgate template configurations labels Aug 10, 2026

mwendigg enabled auto-merge August 10, 2026 04:26

mwendigg merged commit 286b0d9 into dreadnode:main Aug 10, 2026
19 of 30 checks passed

l50 deleted the chore/sync-upstream-code branch August 10, 2026 04:27

Navigation Menu

Uh oh!

feat: add LLM orchestrator, blue-team investigation pipeline, and benchmark replay system - #441

feat: add LLM orchestrator, blue-team investigation pipeline, and benchmark replay system#441
mwendigg merged 485 commits into
dreadnode:mainfrom
l50:chore/sync-upstream-code

l50 commented Aug 10, 2026

Uh oh!

Uh oh!

codecov Bot commented Aug 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

l50 commented Aug 10, 2026

Uh oh!

Uh oh!

codecov Bot commented Aug 10, 2026

Codecov Report

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants