AI Sentry - Unified AI/LLM Security Scanner for Static Code Analysis + Live Model Testing
Project description
AI Security CLI
A unified command-line tool for AI/LLM security scanning and testing. Combines static code analysis with live model testing to provide comprehensive security assessment for AI applications.
Website: aisentry.co
Benchmarks
Evaluated against a comprehensive OWASP LLM Top 10 testbed with 73 ground-truth vulnerabilities.
| Metric | AI Security CLI | Semgrep | Bandit |
|---|---|---|---|
| Precision | 68.5% | 83.3% | 58.3% |
| Recall | 50.7% | 6.8% | 38.4% |
| F1 Score | 58.3% | 12.7% | 46.3% |
LLM-Specific Coverage (patterns generic tools can't detect):
| Category | AI-Sec F1 | Semgrep | Bandit |
|---|---|---|---|
| LLM01: Prompt Injection | 60.0% | 0% | 15.4% |
| LLM04: Model DoS | 80.0% | 0% | 0% |
| LLM06: Sensitive Info | 62.5% | 0% | 0% |
| LLM10: Model Theft | 44.4% | 0% | 0% |
General Patterns (where generic tools excel):
| Category | AI-Sec F1 | Semgrep | Bandit |
|---|---|---|---|
| LLM02: Insecure Output | 35.3% | 42.9% | 81.8% |
| LLM07: Insecure Plugin | 71.4% | 25.0% | 83.3% |
Note: Semgrep/Bandit are general-purpose SAST tools not designed for LLM patterns. Use AI Security CLI + Bandit together for comprehensive coverage. See llm-sec-eval for methodology and limitations.
Features
- Static Code Analysis: Scan Python codebases for OWASP LLM Top 10 vulnerabilities
- Security Posture Audit: Auto-detect security controls and generate maturity scores across 10 categories (61 controls)
- Remote Repository Scanning: Scan GitHub, GitLab, and Bitbucket repositories directly via URL
- Interactive HTML Reports: Modern reports with tabbed interface, dark mode, severity filtering, and pagination
- SARIF Output: CI/CD integration with GitHub Code Scanning, Azure DevOps, VS Code, and more
- Configurable: YAML config files, environment variables, per-category thresholds, test file handling
- 4-Factor Confidence Scoring: Advanced confidence calculation for accurate vulnerability assessment
Live Model Testing
For live/runtime testing of LLM models (prompt injection, jailbreaks, etc.), we recommend Garak - a comprehensive LLM vulnerability scanner by NVIDIA.
# Install Garak
pip install garak
# Run probes against a model
garak --model_type openai --model_name gpt-4 --probes all
AI Security CLI focuses on static code analysis - finding vulnerabilities in your source code before deployment. Garak complements this by testing the runtime behavior of deployed models.
Installation
# Basic installation
pip install aisentry
# With cloud provider support
pip install aisentry[cloud]
# Development installation
pip install aisentry[dev]
# Full installation with all features
pip install aisentry[all]
Configuration
Config File (.ai-security.yaml)
Create a .ai-security.yaml file in your project root:
# Scan mode: recall (high sensitivity) or strict (higher thresholds)
mode: recall
# Deduplication: exact (merge duplicates) or off
dedup: exact
# Directories to exclude
exclude_dirs:
- vendor
- third_party
- node_modules
# Test file handling
exclude_tests: false
demote_tests: true
test_confidence_penalty: 0.25
# Per-category confidence thresholds
thresholds:
LLM01: 0.70
LLM02: 0.70
LLM05: 0.80
LLM06: 0.75
# Global threshold (used if category not specified)
global_threshold: 0.70
Environment Variables
| Variable | Description | Example |
|---|---|---|
AISEC_MODE |
Scan mode | recall or strict |
AISEC_DEDUP |
Deduplication | exact or off |
AISEC_EXCLUDE_DIRS |
Comma-separated dirs | vendor,third_party |
AISEC_THRESHOLD |
Global threshold | 0.70 |
AISEC_THRESHOLD_LLM01 |
Per-category threshold | 0.80 |
Precedence: CLI flags > Environment variables > .ai-security.yaml > Defaults
Quick Start
# Static code analysis (local)
aisentry scan ./my_project
# Static code analysis (remote GitHub repository)
aisentry scan https://github.com/langchain-ai/langchain
# Generate HTML report with Security Posture audit (default)
aisentry scan ./my_project -o html -f security_report.html
# Security posture audit only
aisentry audit ./my_project
# Live model testing
export OPENAI_API_KEY=sk-...
aisentry test -p openai -m gpt-4 --mode quick
HTML Report Features
The HTML reports include a modern, interactive interface:
- Tabbed Interface: Switch between Vulnerabilities and Security Posture views
- Dark Mode: Toggle between light and dark themes (persists in browser)
- Severity Filtering: Click severity buttons to filter by Critical, High, Medium, Low
- Pagination: "Show More" button loads items in batches of 10
- Combined Scoring: See both vulnerability score and security posture score
- Hover Effects: Cards and items highlight on hover for better UX
Architecture
High-Level Overview
┌──────────────────────────────────────────────────────────────────────────────────────┐
│ AI SECURITY CLI │
├──────────────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐ │
│ │ scan command │ │ audit command │ │ test command │ │
│ └───────┬────────┘ └───────┬────────┘ └───────┬────────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌───────────────────┐ ┌───────────────────┐ ┌───────────────────┐ │
│ │ STATIC ANALYSIS │ │ SECURITY AUDIT │ │ LIVE TESTING │ │
│ │ │ │ │ │ │ │
│ │ • AST Parser │ │ • 61 Controls │ │ • 7 LLM Providers │ │
│ │ • 10 OWASP Detect │ │ • 10 Categories │ │ • 11 Detectors │ │
│ │ • 7 Scorers │ │ • Maturity Score │ │ • 4-Factor Conf. │ │
│ └─────────┬─────────┘ └─────────┬─────────┘ └─────────┬─────────┘ │
│ │ │ │ │
│ └─────────────────────┼─────────────────────┘ │
│ ▼ │
│ ┌──────────────────────────────┐ │
│ │ REPORT GENERATION │ │
│ │ JSON | HTML | SARIF | Text │ │
│ │ │ │
│ │ HTML Features: │ │
│ │ • Tabbed Interface │ │
│ │ • Dark Mode Toggle │ │
│ │ • Severity Filtering │ │
│ │ • Pagination │ │
│ └──────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────────────────────────┘
Static Analysis Flow
┌──────────────────────────────────────────────────────────────────────────────────┐
│ STATIC ANALYSIS PIPELINE │
└──────────────────────────────────────────────────────────────────────────────────┘
┌─────────┐ ┌─────────────┐ ┌────────────────────────────────────────┐
│ Python │ │ AST Parser │ │ 10 OWASP DETECTORS │
│ Code │─────▶│ & Pattern │─────▶│ │
│ (.py) │ │ Extractor │ │ ┌──────────┐ ┌──────────┐ ┌────────┐ │
└─────────┘ └─────────────┘ │ │ LLM01 │ │ LLM02 │ │ LLM03 │ │
│ │ Prompt │ │ Insecure │ │Training│ │
│ │ Injection│ │ Output │ │Poison │ │
│ └──────────┘ └──────────┘ └────────┘ │
│ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│ │ LLM04 │ │ LLM05 │ │ LLM06 │ │
│ │Model DoS │ │ Supply │ │Secrets │ │
│ │ │ │ Chain │ │ │ │
│ └──────────┘ └──────────┘ └────────┘ │
│ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│ │ LLM07 │ │ LLM08 │ │ LLM09 │ │
│ │ Insecure │ │Excessive │ │ Over │ │
│ │ Plugin │ │ Agency │ │reliance│ │
│ └──────────┘ └──────────┘ └────────┘ │
│ ┌──────────┐ │
│ │ LLM10 │ │
│ │ Model │ │
│ │ Theft │ │
│ └──────────┘ │
└───────────────────┬────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────────────────────┐
│ 7 SECURITY SCORERS │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Prompt │ │ Model │ │ Data │ │Hallucin- │ │ Ethical │ │
│ │ Security │ │ Security │ │ Privacy │ │ ation │ │ AI │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ ┌──────────┐ ┌──────────┐ │
│ │Governance│ │ OWASP │ │
│ │ │ │ Score │ │
│ └──────────┘ └──────────┘ │
└───────────────────────────────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────┐
│ SCAN RESULT │
│ • Findings • Category Scores│
│ • Overall Score • Confidence │
└─────────────────────────────────┘
Live Testing Flow
┌──────────────────────────────────────────────────────────────────────────────────┐
│ LIVE TESTING PIPELINE │
└──────────────────────────────────────────────────────────────────────────────────┘
┌────────────────────────────────────────────────────────────────────────────────┐
│ 7 LLM PROVIDERS │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌───────┐ ┌─────┐│
│ │ OpenAI │ │Anthropic│ │ AWS │ │ Google │ │ Azure │ │Ollama │ │Cust-││
│ │ │ │ │ │ Bedrock │ │ Vertex │ │ OpenAI │ │(local)│ │ om ││
│ └────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘ └───┬───┘ └──┬──┘│
└───────┴──────────┴──────────┴──────────┴──────────┴─────────┴────────┴──────┘
│
▼
┌──────────────────────────┐
│ BASELINE QUERIES │
└────────────┬─────────────┘
│
▼
┌────────────────────────────────────────────────────────────────────────────────┐
│ 11 LIVE DETECTORS │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐│
│ │ Prompt │ │Jailbreak │ │ Data │ │ Halluc- │ │ DoS │ │ Bias ││
│ │Injection │ │ │ │ Leakage │ │ ination │ │ │ │Detection ││
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ └──────────┘ └──────────┘│
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Model │ │Adversar- │ │ Output │ │ Supply │ │Behavioral│ │
│ │Extraction│ │ ial │ │ Manip. │ │ Chain │ │ Anomaly │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
└───────────────────────────────────────────┬────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────────────────────┐
│ 4-FACTOR CONFIDENCE CALCULATION │
│ │
│ Response Analysis (30%) + Detector Logic (35%) + │
│ Evidence Quality (25%) + Severity Factor (10%) = Confidence Score │
└───────────────────────────────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────┐
│ TEST RESULT │
│ • Vulnerabilities • Score │
│ • Tests Passed • Confidence │
└─────────────────────────────────┘
Component Architecture
┌─────────────────────────────────────────────────────────────────────────────────┐
│ aisentry package │
├─────────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────────────────────────────────────────────────────────────────┐ │
│ │ CLI LAYER (cli.py) │ │
│ │ scan command ─────────────────────────── test command │ │
│ └─────────┬───────────────────────────────────────────┬───────────────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌──────────────────────────┐ ┌──────────────────────────┐ │
│ │ scanner.py │ │ tester.py │ │
│ └────────────┬─────────────┘ └────────────┬─────────────┘ │
│ │ │ │
│ ┌────────┴────────┐ ┌─────────┴─────────┐ │
│ ▼ ▼ ▼ ▼ │
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ │
│ │ STATIC │ │ SCORERS │ │ LIVE │ │ PROVIDERS │ │
│ │ DETECTORS │ │ │ │ DETECTORS │ │ │ │
│ │ LLM01-10 │ │ 7 scorers │ │ 11 detects │ │ 7 providers│ │
│ └────────────┘ └────────────┘ └────────────┘ └────────────┘ │
│ │
│ ┌─────────────────────────────────────────────────────────────────────────┐ │
│ │ REPORTERS: base | json | html | sarif │ │
│ └─────────────────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────────────────┐ │
│ │ MODELS: finding.py | vulnerability.py | result.py │ │
│ └─────────────────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────────────────┐ │
│ │ UTILS: markov_chain | entropy | scoring | statistical │ │
│ └─────────────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────────┘
CLI Commands
Static Code Analysis (scan)
Scan Python code for OWASP LLM Top 10 vulnerabilities. Supports local files/directories and remote Git repositories.
aisentry scan <path> [OPTIONS]
Path Options:
| Path Type | Example |
|---|---|
| Local file | ./app.py |
| Local directory | ./my_project |
| GitHub URL | https://github.com/user/repo |
| GitLab URL | https://gitlab.com/user/repo |
| Bitbucket URL | https://bitbucket.org/user/repo |
Options:
| Option | Description | Default |
|---|---|---|
-o, --output |
Output format: text, json, html, sarif | text |
-f, --output-file |
Write output to file | - |
-s, --severity |
Minimum severity: critical, high, medium, low, info | info |
-c, --confidence |
Minimum confidence threshold (0.0-1.0) | 0.7 |
--category |
Filter by OWASP category (LLM01-LLM10) | all |
--audit/--no-audit |
Include security posture audit in HTML reports | true |
--config |
Path to .ai-security.yaml config file | auto-detect |
--mode |
Scan mode: recall (sensitive) or strict (precise) | recall |
--dedup |
Deduplication: exact (merge) or off | exact |
--exclude-dir |
Directories to exclude (repeatable) | - |
--exclude-tests |
Skip test files entirely | false |
--demote-tests |
Reduce confidence for test file findings | true |
-v, --verbose |
Enable verbose output | false |
Examples:
# Scan a local project directory
aisentry scan ./my_llm_app
# Scan with JSON output
aisentry scan ./app.py -o json -f results.json
# Scan for high severity issues only
aisentry scan ./project -s high
# Scan specific OWASP categories
aisentry scan ./project --category LLM01 --category LLM02
# Generate HTML report
aisentry scan ./project -o html -f security_report.html
# Scan a GitHub repository directly
aisentry scan https://github.com/langchain-ai/langchain
# Generate HTML without security posture audit
aisentry scan ./project -o html --no-audit -f vuln-only.html
Security Posture Audit (audit)
Evaluate security controls and maturity level of your codebase. Detects 61 security controls across 10 categories.
aisentry audit <path> [OPTIONS]
Options:
| Option | Description | Default |
|---|---|---|
-o, --output |
Output format: text, json, html | text |
-f, --output-file |
Write output to file | - |
-v, --verbose |
Enable verbose output | false |
Security Control Categories:
| Category | Controls | Description |
|---|---|---|
| Prompt Security | 8 | Input validation, sanitization, injection prevention, red teaming |
| Model Security | 8 | Rate limiting, access controls, model protection, differential privacy |
| Data Privacy | 8 | PII detection, encryption, data anonymization, GDPR compliance |
| OWASP LLM Top 10 | 10 | Coverage of OWASP LLM security controls |
| Blue Team Operations | 7 | Logging, monitoring, alerting, drift detection |
| Governance | 5 | Compliance, documentation, audit trails |
| Supply Chain | 3 | Dependency scanning, model provenance, integrity verification |
| Hallucination Mitigation | 5 | RAG implementation, confidence scoring, fact checking |
| Ethical AI & Bias | 4 | Fairness metrics, explainability, bias testing, model cards |
| Incident Response | 3 | Monitoring integration, audit logging, rollback capability |
Maturity Levels:
| Level | Score | Description |
|---|---|---|
| Initial | 0-20 | No formal security controls |
| Developing | 21-40 | Basic controls being implemented |
| Defined | 41-60 | Documented security processes |
| Managed | 61-80 | Measured and controlled security |
| Optimizing | 81-100 | Continuous security improvement |
Examples:
# Audit a local project
aisentry audit ./my_project
# Generate HTML audit report
aisentry audit ./project -o html -f audit-report.html
# Audit a GitHub repository
aisentry audit https://github.com/user/repo -o json
Live Model Testing (test)
Test live LLM models for security vulnerabilities.
aisentry test [OPTIONS]
Options:
| Option | Description | Default |
|---|---|---|
-p, --provider |
LLM provider (required) | - |
-m, --model |
Model name (required) | - |
-e, --endpoint |
Custom endpoint URL | - |
-t, --tests |
Specific tests to run | all |
--mode |
Testing depth: quick, standard, comprehensive | standard |
-o, --output |
Output format: text, json, html, sarif | text |
-f, --output-file |
Write output to file | - |
--timeout |
Timeout per test in seconds | 30 |
-v, --verbose |
Enable verbose output | false |
Supported Providers:
| Provider | Environment Variables |
|---|---|
openai |
OPENAI_API_KEY |
anthropic |
ANTHROPIC_API_KEY |
bedrock |
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY |
vertex |
GOOGLE_APPLICATION_CREDENTIALS |
azure |
AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT |
ollama |
None (local) |
custom |
CUSTOM_API_KEY (optional) |
Examples:
# Quick test with OpenAI
export OPENAI_API_KEY=sk-...
aisentry test -p openai -m gpt-4 --mode quick
# Comprehensive test with Anthropic
export ANTHROPIC_API_KEY=...
aisentry test -p anthropic -m claude-3-opus --mode comprehensive
# Test specific vulnerabilities
aisentry test -p openai -m gpt-4 -t prompt-injection -t jailbreak
# Test with Ollama (local)
aisentry test -p ollama -m llama2 --mode standard
OWASP LLM Top 10 Coverage
Static Analysis Detectors
| ID | Vulnerability | Description |
|---|---|---|
| LLM01 | Prompt Injection | Detects unsanitized user input in prompts |
| LLM02 | Insecure Output Handling | Identifies unvalidated LLM output |
| LLM03 | Training Data Poisoning | Finds unsafe data loading |
| LLM04 | Model Denial of Service | Detects missing rate limiting |
| LLM05 | Supply Chain Vulnerabilities | Identifies unsafe model loading |
| LLM06 | Sensitive Information Disclosure | Finds hardcoded secrets |
| LLM07 | Insecure Plugin Design | Detects unsafe plugin loading |
| LLM08 | Excessive Agency | Identifies autonomous actions |
| LLM09 | Overreliance | Finds missing output validation |
| LLM10 | Model Theft | Detects exposed model artifacts |
Live Testing Detectors
| ID | Detector | Description |
|---|---|---|
| PI | Prompt Injection | Tests for injection vulnerabilities |
| JB | Jailbreak | Tests for instruction bypass attacks |
| DL | Data Leakage | Tests for PII exposure |
| HAL | Hallucination | Tests for factual accuracy |
| DOS | Denial of Service | Tests for resource exhaustion |
| BIAS | Bias Detection | Tests for demographic bias |
| ME | Model Extraction | Tests for architecture disclosure |
| ADV | Adversarial Inputs | Tests for encoding attacks |
| OM | Output Manipulation | Tests for response injection |
| SC | Supply Chain | Tests for unsafe code generation |
| BA | Behavioral Anomaly | Tests for unexpected behavior |
Output Formats
- Text: Human-readable terminal output
- JSON: Machine-readable format for CI/CD
- HTML: Interactive reports with filtering
- SARIF: GitHub Code Scanning, Azure DevOps, VS Code integration
Integration
GitHub Actions
name: AI Security Scan
on: [push, pull_request]
jobs:
security-scan:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v3
- uses: actions/setup-python@v4
with:
python-version: '3.11'
- run: pip install aisentry
- run: aisentry scan . -o sarif -f results.sarif
- uses: github/codeql-action/upload-sarif@v2
with:
sarif_file: results.sarif
Pre-commit Hook
repos:
- repo: local
hooks:
- id: ai-security-scan
name: AI Security Scan
entry: aisentry scan
language: system
types: [python]
args: ['-s', 'high']
Development
git clone https://github.com/deosha/aisentry.git
cd aisentry
pip install -e ".[dev]"
pytest tests/ -v --cov=aisentry
License
MIT License - see LICENSE for details.
Links
- Website: aisentry.co
- GitHub: github.com/deosha/aisentry
- PyPI: pypi.org/project/aisentry
- Issues: Report bugs
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
aisentry-1.0.0b1.tar.gz
(289.2 kB
view details)
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
aisentry-1.0.0b1-py3-none-any.whl
(324.1 kB
view details)
File details
Details for the file aisentry-1.0.0b1.tar.gz.
File metadata
- Download URL: aisentry-1.0.0b1.tar.gz
- Upload date:
- Size: 289.2 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.13.6
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
57be0e956e063b91d6fe1c0e2794712a1fc0ec097804bf7a37fb074e7f30d9ba
|
|
| MD5 |
1d4835ec9267a659825a8d7a312ffbd0
|
|
| BLAKE2b-256 |
61339a9dee4de60ad2167cfaa36aeb1f7800f7c74d30d43fd90b12e52517bd92
|
File details
Details for the file aisentry-1.0.0b1-py3-none-any.whl.
File metadata
- Download URL: aisentry-1.0.0b1-py3-none-any.whl
- Upload date:
- Size: 324.1 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.13.6
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
e3855d7a5d36018a694b7c2a49ab1a955d638149c738cf0e3148dddd160c3540
|
|
| MD5 |
d981d39f7af8923cc41bb8f702050a3d
|
|
| BLAKE2b-256 |
ef8636b06f2ca56b878fee29388fbda13c2065354dee4dd63f8b0cea3d46f464
|
