community-ai-audit 0.5.0

Community-driven AI security audit tool using interpretability techniques

Community AI Security Audit Tool

A plug-and-play, community-driven AI security audit framework. Scan any model (local, HF Hub, OpenAI, Anthropic, AWS Bedrock, Ollama) for vulnerabilities, interpret decisions, and push findings to your SIEM (Splunk, Elastic, Datadog, Sentinel).

Features

• Model Adapters: HuggingFace, OpenAI, Anthropic, AWS Bedrock, Local (PyTorch/ONNX/SafeTensors), Ollama
• Vulnerability Scanners: Backdoor/Trojan detection (activation clustering), Adversarial robustness (FGSM/PGD)
• Interpretability: Integrated Gradients, LIME
• SIEM Connectors: Splunk HEC, Elastic/Elasticsearch, Datadog Logs, Microsoft Sentinel
• Reporting: Markdown, JSON, HTML
• CLI: discover, scan, interpret, audit commands
• Extensible: Plugin system for custom adapters, scanners, interpreters, connectors

Quickstart

# Install
pip install community-ai-audit[hf]  # with HuggingFace support

# Discover available plugins
community-ai-audit discover

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

# Full audit (scan + interpret) with SIEM push
community-ai-audit audit meta-llama/Llama-3-8B-Instruct \
  --provider huggingface \
  --profile standard \
  --scanners backdoor adversarial \
  --interpreters integrated-gradients lime \
  --connectors splunk elastic \
  --config config/my_connectors.yaml

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      AuditEngine                                │
│  (orchestrates: load → scan → interpret → report → push)       │
└─────────────────────────────────────────────────────────────────┘
         │                    │                    │
         ▼                    ▼                    ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│   Adapters      │ │   Scanners      │ │  Interpreters   │
│  (load_model)   │ │  (scan model)   │ │ (explain model) │
├─────────────────┤ ├─────────────────┤ ├─────────────────┤
│ • HuggingFace   │ │ • Backdoor      │ │ • Integrated    │
│ • OpenAI        │ │ • Adversarial   │ │   Gradients     │
│ • Anthropic     │ │ • (custom)      │ │ • LIME          │
│ • AWS Bedrock   │ │                 │ │ • (custom)      │
│ • Local         │ │                 │ │                 │
│ • Ollama        │ │                 │ │                 │
└─────────────────┘ └─────────────────┘ └─────────────────┘
         │                    │                    │
         └────────────────────┼────────────────────┘
                              ▼
                   ┌─────────────────────┐
                   │   Reporters         │
                   │ (markdown, json,    │
                   │  html)              │
                   └─────────────────────┘
                              │
                              ▼
                   ┌─────────────────────┐
                   │  SIEM Connectors    │
                   │ (push findings)     │
                   ├─────────────────────┤
                   │ • Splunk HEC        │
                   │ • Elastic           │
                   │ • Datadog           │
                   │ • Sentinel          │
                   │ • (custom)          │
                   └─────────────────────┘

Provider Matrix

Provider	Text	Image	Multimodal	Embedding	Auth
HuggingFace	✅	✅	✅	✅	HF_TOKEN
OpenAI	✅	✅	✅	✅	OPENAI_API_KEY
Anthropic	✅	❌	❌	❌	ANTHROPIC_API_KEY
AWS Bedrock	✅	✅	✅	✅	AWS creds
Local (PyTorch)	✅	✅	✅	✅	None
Ollama	✅	❌	❌	❌	Local server

Installation

# Core only
pip install community-ai-audit

# With HuggingFace transformers
pip install community-ai-audit[hf]

# With TensorFlow support
pip install community-ai-audit[tf]

# All optional dependencies
pip install community-ai-audit[all]

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

Configuration

Create config/my_config.yaml:

model:
  device: auto
  dtype: auto

scanners:
  backdoor:
    enabled: true
    num_clusters: 5
    activation_threshold: 0.85
  adversarial:
    enabled: true
    epsilon: 0.1
    pgd_steps: 10

connectors:
  splunk:
    hec_url: https://splunk.example.com:8088
    hec_token: ${SPLUNK_HEC_TOKEN}
    index: security
  elastic:
    url: https://es.example.com:9243
    api_key: ${ELASTICSEARCH_API_KEY}

Use with --config config/my_config.yaml.

CLI Commands

Discover

community-ai-audit discover
community-ai-audit discover --format json

Scan

# Quick scan with defaults
community-ai-audit scan distilgpt2 --provider huggingface

# Custom profile and scanners
community-ai-audit scan model.pt --provider local \
  --profile deep \
  --scanners backdoor adversarial \
  --input-shape '[32, 768]' \
  --output markdown --save report.md

Interpret

community-ai-audit interpret distilgpt2 --provider huggingface \
  --interpreters integrated-gradients lime \
  --input "The model should classify this as positive."

Audit (Full Pipeline)

community-ai-audit audit meta-llama/Llama-3-8B-Instruct \
  --provider huggingface \
  --profile standard \
  --scanners backdoor adversarial \
  --interpreters integrated-gradients \
  --connectors splunk elastic \
  --config config/my_connectors.yaml

Extending

See Plugin Guide for:

• Adding a Model Adapter
• Adding a SIEM Connector
• Adding a Scanner
• Adding an Interpreter

Known Limitations

• Scanners require white-box access (gradients/activations) — work best with provider=local or HuggingFace local models
• Text model adversarial attacks need embedding-space perturbations (token IDs are discrete)
• Integrated Gradients on text requires access to embedding layer
• Large model audits can be slow — batch mode coming in v0.2.0
• TensorFlow support planned

Contributing

See CONTRIBUTING.md and Plugin Guide.

License

MIT License — see LICENSE.