Hiding a VM should be harder than finding one.
This is a silly little script that spiraled completely out of control.
What started as curiosity about VM detection after watching a YouTube video about a VM hiding tool turned into a multi-vector detection engine with tiered confidence scoring, direct CPUID access, parallel artifact gathering, and anti-analysis heuristics.
The question I had was simple: "How can one still detect a VM even if everything gets obfuscated or scrambled?"
Important
This project is foremost a learning exercise. It is NOT intended for malicious use. It is for educational/personal use only (or ethical red team tooling, maybe. But idk if this is good enough for that yet lol).
PyVM uses a three-phase detection pipeline:
┌────────────────────────────────────────────────────────┐
│ DETECTION PIPELINE │
├────────────────────────────────────────────────────────┤
│ │
│ PHASE 1: GATHER (parallel, ~2s) │
│ ├── 10 independent gatherers (ThreadPoolExecutor) │
│ │ PCI · ACPI · MAC · Disk · GPU · FS · Quirks · │
│ │ Cloud · Uptime · Container │
│ │ │
│ ├── 9 sequential gatherers (dependency-ordered) │
│ │ CPUID · BIOS · Processes · Interrupts · Entropy · │
│ │ Topology · Timing · Memory · Nested Virt │
│ │ │
│ └── Normalize & deduplicate │
│ │
│ PHASE 2: SANDBOX CHECKS │
│ ├── Process scanning (30+ signatures) │
│ ├── Environment variable inspection │
│ ├── Username/hostname heuristics │
│ ├── Sleep timing validation │
│ └── Debugger/tracer detection │
│ │
│ PHASE 3: TIERED SCORING │
│ ├── T1 HARD (1.0×) —> PCI, CPUID, MAC, ACPI │
│ ├── T2 MEDIUM (0.6×) —> BIOS, Disk, Processes, GPU │
│ ├── T3 SOFT (0.3×) —> Entropy, Timing, Topology │
│ └── Sandbox penalties + classification │
│ │
└────────────────────────────────────────────────────────┘
Not all evidence is equal. PyVM weighs signals by how hard they are to spoof:
| Tier | Weight | Signals | Spoof Difficulty |
|---|---|---|---|
| T1: Hard | 1.0× | PCI vendor/device IDs, direct CPUID signatures, MAC OUI prefixes, ACPI tables | Near-impossible (firmware-level) |
| T2: Medium | 0.6× | BIOS/SMBIOS strings, disk vendors, VM processes, virtual GPUs | Moderate (requires system editing) |
| T3: Soft | 0.3× | Entropy timing, instruction latency, CPU topology, uptime, RAM sizing | Easiest to manipulate |
A default unstealthed VM gets caught by multiple T1 signals simultaneously (easily hitting 100%). Even partially stealthed VMs that hide CPUID and spoof BIOS still get caught by PCI vendors and MAC prefixes.
More to come
Hardware & Firmware Signatures
| Vector | Method | Tier |
|---|---|---|
| PCI vendor/device IDs | sysfs + lspci (Linux), Registry (Windows) | T1 |
| CPUID hypervisor signatures | Native C library (direct CPUID instruction) | T1 |
| CPUID leaf 0x40000000+ | Extended hypervisor vendor strings | T1 |
| CPUID ECX bit 31 | Hypervisor present flag | T3 |
| ACPI table signatures | /sys/firmware/acpi/tables/ | T1 |
| MAC OUI prefixes | psutil network interfaces | T1 |
| BIOS/SMBIOS vendor | dmidecode (sudo-aware) + sysfs fallback | T2 |
| Disk vendor strings | /sys/block + /dev/disk/by-id/ | T2 |
| GPU adapter identification | lspci -v / wmic | T2 |
| Filesystem artifacts | /dev/vda, /proc/xen, VBoxGuest.sys, etc. | T2 |
Behavioral Heuristics
| Vector | Method | What It Detects |
|---|---|---|
| Entropy timing variance | os.urandom(16384) × 8500 samples |
VM-exit latency spikes during CSPRNG refilling |
| Instruction timing | sum(range(50)) × 25 batches of 200 |
Privileged operation overhead |
| Interrupt jitter | time.sleep(0) × 1000 samples |
Abnormally tight scheduling (emulation artifact) |
| CPU topology ratio | Logical/physical core ratio | Suspicious thread-core configurations |
| Memory sizing patterns | Power-of-two allocation detection | Common VM defaults (2/4/8/16/32 GB) |
| Uptime analysis | psutil boot time | Recently booted sandboxes |
| Nested virtualization | VT-x/SVM flags + VirtIO + timing variance | VM-inside-VM scenarios |
Environment Classification
| Vector | Method |
|---|---|
| Container detection | /.dockerenv + cgroup ID hashes + docker.sock (multi-signal required) |
| Cloud metadata probing | AWS / Azure / GCP / Alibaba / DigitalOcean endpoints |
| Sandbox/anti-analysis | 30+ process signatures, env vars, debuggers, timing anomalies |
| Sleep timing validation | Detects accelerated/skipped sleeps (sandbox artifact) |
- Python 3.12+
psutil(required)beaupy(optional — enables the interactive TUI)requests(optional — enables cloud metadata probing)gcc(optional — for compiling the direct CPUID library)
# Clone the repo
git clone https://github.com/therealOri/PyVM.git
cd PyVM
# Install Python dependencies
pip install psutil beaupy requests
# (Optional) Compile the direct CPUID library for unforgeable detection
(Build and setup libcpuid)
# Run the TUI
python vm_ui.py
# Or run headless
python vm_core.pypython vm_ui.pyProvides a beaupy-powered menu with three scan modes:
| Mode | Description |
|---|---|
| Quick Scan | Fast scan with score breakdown and artifacts |
| Detailed Scan | Adds tier-by-tier detection reasoning (T1/T2/T3) |
| Debug Mode | Full raw JSON dump of every artifact and signal |
from vm_core import Detector, set_prompt_callbacks
detector = Detector()
# Full scan with explanations
result = detector.detect(parallel=True, explain=True)
print(f"Classification: {result['classification']}")
print(f"Confidence: {result['confidence']}%")
print(f"Best Guess: {result['best_guess']}")
# Access raw scores
for vm, score in sorted(result['scores'].items(), key=lambda x: -x[1]):
if score > 0:
print(f" {vm}: {score}%")
# Check anti-analysis environment
if result['anti_analysis']['detected']:
print("⚠️ Sandbox detected!")
# Access full artifact data
artifacts = result['artifacts']
print(f"CPUID Signature: {artifacts['cpuid_signature']}")
print(f"Entropy median: {artifacts['entropy_behavior']['median_ns']:,}ns")| Hypervisor | Default (No Stealth) | Partial Stealth¹ | Deep Stealth² |
|---|---|---|---|
| QEMU/KVM | ✅ 100% | ✅ 90%+ | |
| VirtualBox | ✅ 100% | ✅ 90%+ | |
| VMware | ✅ 100% | ✅ 90%+ | |
| Hyper-V | ✅ 90%+ | ✅ 85%+ | |
| Xen | ✅ 85%+ | ✅ 80%+ |
¹ Hides CPUID + spoofs BIOS/MAC ² Above + spoofs PCI vendors (requires kernel-level hypervisor modifications)
Reality check: No guest-side detection tool catches a fully stealthed hypervisor with all signatures scrubbed. PyVM handles everything up to deep PCI spoofing, which puts it in the top tier of detection tools.
| Feature | Linux | Windows |
|---|---|---|
| PCI vendor/device scanning | ✅ | ✅ |
| Direct CPUID access | ✅ (.so) | ✅ (.dll) |
| ACPI table parsing | ✅ | ❌ |
| BIOS/SMBIOS via dmidecode | ✅ | ✅ (wmic) |
| Entropy timing heuristics | ✅ | ✅ |
| Instruction timing | ✅ | ✅ |
| Container detection | ✅ | ❌ |
| Cloud metadata probing | ✅ | ✅ |
| Sandbox/anti-analysis | ✅ | ✅ |
| TUI (beaupy) | ✅ | ✅ |
Short-term (v2.1)
- RDTSC timing (sub-nanosecond via timestamp counter)
- CPUID leaf 0x40000003 (KVM paravirt feature flags)
- MSR probing via
/dev/cpu/0/msr - Disk serial number pattern analysis (QM00001, VB00001)
- DMI table version checking
- Cache topology analysis (
lscpu --caches) - Cross-validation scoring (inconsistency detection between vectors)
Long-term (v3.0)
- Machine learning classifier on timing distributions
- TPM remote attestation
- Windows registry deep-scan
- TCP/IP stack fingerprinting
- Side-channel attacks (Flush+Reload, Prime+Probe)
- Spoofing detection scoring (detect the spoofer, not just the VM)
If this project helped you learn something or saved you some time, consider supporting me and helping me get more coffee:
💵 Donate via Cash App
v2.0 —> Current
- Rewritten scoring engine with tiered confidence weighting (T1/T2/T3)
- Direct CPUID access via native C library
- Parallel artifact gathering (ThreadPoolExecutor)
- Cloud metadata endpoint probing (AWS/Azure/GCP/Alibaba/DigitalOcean)
- Container detection with multi-signal requirement
- Sudo-aware privilege handling with UI callback hooks
- Anti-analysis sandbox heuristics (30+ signatures)
- Entropy timing analysis (VM-exit outlier detection)
- Instruction timing analysis (privileged op overhead)
- Nested virtualization detection
- Beaupy-powered TUI with animated spinner
- False-positive elimination on bare metal (calibrated baselines)
v1.0 —> Initial Release
- Basic PCI vendor matching
- CPUID signature checking via /proc/cpuinfo
- MAC OUI prefix detection
- BIOS vendor keyword matching
- Simple additive scoring (no tiers)
Made with curiosity and too much caffeine
Got feedback or ideas? Open an issue or reach out! ^-^