community-ai-audit 0.6.0

Community-driven AI security audit tool using interpretability techniques

🛡️ Community AI Audit

Enterprise-Grade AI Security Auditing · Open Source · Community-Driven

Community AI Audit is a unified security auditing platform for AI/ML models. It provides vulnerability scanning, red team attack simulations, mechanistic interpretability analysis, alignment auditing, and a unified 7-dimension scoring engine — all from a single CLI.

---

🔍 What You Can Do

Use Case	What It Solves
🛡️ Vulnerability Scanning	Detect adversarial susceptibility, backdoors, prompt injection, data extraction, toxicity, watermark detectability
⚔️ Red Team Testing	Simulate jailbreak, multi-turn, obfuscation, roleplay, and tool exploitation attacks
🧠 Mechanistic Interpretability	Probe representations, attention patterns, feature attribution, and layer behavior
🎯 Alignment Auditing	Measure sycophancy, preference drift, value alignment, and objective robustness
📊 Unified Scoring	Aggregate 7 security dimensions into a single risk score with configurable weights
📈 Trend Tracking	Monitor score evolution across time and detect regressions
📡 SIEM Integration	Push findings to Splunk, Elastic, Datadog, Sentinel, and 9+ other platforms

---

⚡ Quickstart

pip install community-ai-audit

# Discover available plugins
community-ai-audit discover

# Scan a model
community-ai-audit scan distilgpt2 --provider huggingface --profile quick

# Full audit with SIEM push
community-ai-audit audit meta-llama/Llama-3-8B-Instruct \
  --provider huggingface --profile standard \
  --connectors splunk elastic

# Red team attack simulation
community-ai-audit redteam gpt-4 --provider openai

# Alignment auditing
community-ai-audit alignment claude-3-opus --provider anthropic

# Compute unified 7-dimension score
community-ai-audit audit-score \
  --scan scan_results.json \
  --redteam redteam_results.json \
  --alignment alignment_results.json

---

🧩 Capabilities

Model Support — 9 Adapters

Provider	Adapter	Auto-Detect
HuggingFace	huggingface	*/* or llama*
OpenAI	openai	gpt-*, o1*, o3*
Anthropic	anthropic	claude-*
AWS Bedrock	aws_bedrock	—
Local (PyTorch/TF/ONNX)	local	Path/URI/*.pt/*.onnx
Ollama	ollama	name:tag (no /)
Replicate	replicate	—
VertexAI	vertexai	—
Groq	groq	—

Security Scanning — 7 Scanners

Scanner	What It Detects	Technique
adversarial	FGSM/PGD perturbation susceptibility	Gradient-based attacks
backdoor	Triggered malicious behavior	Activation clustering
prompt_injection	Injection vulnerabilities	Heuristic pattern matching
data_extraction	Training data / secret extraction	Response entropy analysis
toxicity	Toxic / biased outputs	Keyword + classifier scoring
watermark	Watermark detectability	Statistical pattern analysis
dsl	User-defined rules	YAML rule engine

Red Team — 5 Attack Scanners

Scanner	Attack Surface	Evaluation
jailbreak	20 known jailbreak prompts	Refusal vs success pattern matching
multi_turn_attack	10 two-turn conversation attacks	Suspicious-keyword breach detection
prompt_obfuscation	10 obfuscated variants (base64, leetspeak)	Harmful-keyword matching
roleplay_attack	15 roleplay scenarios (DAN, character shells)	Refusal vs engagement patterns
tool_exploitation	10 tool-misuse prompts	Exploit-keyword detection

Mechanistic Interpretability — 5 Analyzers

Analyzer	Probes	What It Measures
activation_probes	5 probe inputs	Response quality, SNR estimate
representation_analysis	8 probes, 4 pairs	Jaccard differentiation, vocabulary size
attention_head_analysis	5 syntactic probes	Attention complexity estimate
feature_attribution	5 sentiment inputs	Word-level importance, sentiment match
layer_analysis	3 open-ended probes	Depth estimation, complexity distribution

Alignment Auditing — 4 Scanners

Scanner	Prompts	What It Detects
sycophancy	5 agree + 5 disagree	Stance-sycophancy (rubber-stamping)
preference_drift	5 cores × 3 variants	Sentiment inconsistency across paraphrases
value_alignment	8 probes across 6 values	Refusal of harmful, encouragement of prosocial
objective_robustness	3 objectives × 4 prompts	Refusal-pattern violations per objective

Scoring — 7 Dimensions

┌─────────────────────────────────────────────┐
│           Unified Audit Score                │
├──────────────┬──────────────────────────────┤
│ Security     │   ████████████████░░ 82.0     │
│ Reliability  │   ██████████████░░░░ 72.0     │
│ Compliance   │   ██████████████████ 90.0     │
│ Agent Risk   │   ████████████████░░ 80.0     │
│ Alignment    │   ████████████████░░ 85.0     │
│ Red Team     │   ████████████░░░░░░ 60.0     │
│ Interpretability │ ████████████░░░░░░ 65.0   │
├──────────────┴──────────────────────────────┤
│ Overall: 77.6 (Good)                        │
│ Weights: security=0.2, reliability=0.1, ... │
└─────────────────────────────────────────────┘

Executive Dashboard

Real-time HTML dashboard served via dashboard_v2/server.py:

• 7 color-coded score cards (critical → excellent)
• JSON overlay endpoint for programmatic updates
• Configurable refresh interval
• Responsive CSS grid layout

---

🔧 Installation

# Core (numpy, pyyaml, scikit-learn)
pip install community-ai-audit

# Optional extras
pip install community-ai-audit[torch]      # Torch-based scanners
pip install community-ai-audit[scheduler]  # Cron scheduling
pip install community-ai-audit[hf]         # HuggingFace transformers
pip install community-ai-audit[tf]         # TensorFlow

# Development
git clone https://github.com/anomalyco/community-ai-audit
cd community-ai-audit
pip install -e .[dev]

---

💻 CLI Reference

Command	Description
scan <model> -p <provider>	Run vulnerability scanners
interpret <model> -p <provider>	Run interpretability methods
audit <model> -p <provider>	Full pipeline: scan + interpret + report + push
redteam <model> -p <provider>	Red team attack simulations
mechinterp <model> -p <provider>	Mechanistic interpretability analysis
alignment <model> -p <provider>	Alignment auditing
audit-score	Compute unified 7-dimension score
discover	List all discovered plugins
schedule add/list/remove/run	Manage recurring audits

Exit codes: 0 = ok, 1 = HIGH/MEDIUM findings, 2 = CRITICAL findings.

---

📚 Documentation

Resource	Description
Architecture & Reference	Full component docs, API, CLI, config, deployment
Plugin Guide	Writing custom adapters, scanners, interpreters
Scanner Guide	Details on each vulnerability scanner
Adapter Guide	Details on each model adapter
Connector Guide	SIEM and storage connector details
Red Team	Attack framework and scanner reference
Mech Interp	Analyzer reference and methodology
Alignment	Alignment scanner reference
Scoring Engine	7-dimension scoring details
Dashboard	Executive dashboard server

---

📋 Configuration

cache:
  enabled: true
  max_size: 1000
  ttl_seconds: 3600

scanners:
  adversarial:
    num_samples: 32
    pgd_steps: 10
  backdoor:
    sample_size: 128

connectors:
  splunk:
    url: "${SPLUNK_URL}"
    token: "${SPLUNK_TOKEN}"
  elastic:
    url: "${ELASTIC_URL}"
    api_key: "${ELASTIC_API_KEY}"

Config values can also be set via environment variables: COMMUNITY_AI_AUDIT_CONNECTORS_SPLUNK_URL.

Precedence (lowest → highest): default.yaml → --config PATH → env vars → CLI args.

API Key Safety

1. COMMUNITY_AI_AUDIT_API_KEY env var (recommended)
2. --api-key-file PATH (reads from file, not visible in ps)
3. --api-key VALUE (⚠️ visible in process list)

---

🚀 Deployment

# Docker
docker build -t community-ai-audit .
docker run -v $(pwd)/config:/app/config community-ai-audit scan model.pt -p local

# Docker Compose
docker-compose up -d

# Helm (Kubernetes)
helm install community-ai-audit ./charts/community-ai-audit

# Air-Gapped
./scripts/airgap-bundle.sh   # On connected machine
./scripts/offline-install.sh  # On air-gapped machine

---

🧪 Testing

# All tests (no torch/croniter needed)
pytest tests/

# With coverage
pytest --cov=community_ai_audit tests/

508+ tests covering unit, integration, CLI, connectors, red team, mechanistic interpretability, alignment, trend tracking, and drift analysis.

---

🤝 Contributing

We welcome contributions! See our Plugin Guide to get started writing custom scanners, adapters, or connectors.

---

📄 License

MIT © Anomaly Co.