Security scanner for AI/LLM projects — find leaked API keys, unsafe MCP configs, prompt injection vectors, and vulnerable dependencies before they reach production.

• Secrets Detection — 25+ AI/ML providers covered (OpenAI, Anthropic, Google, HuggingFace, Cohere, Pinecone, LangSmith, AWS, and more), with contextual matching to reduce false positives
• MCP Config Audit — dangerous permissions, hardcoded secrets, insecure HTTP transport, shell execution tools, missing input schemas, spoofable headers
• Prompt Injection Analysis — 12 pattern rules: override instructions, jailbreak keywords, system prompt extraction, unsanitized f-strings, unsafe .format(), Jinja2 SSTI
• Dynamic Endpoint Probing — 24 built-in payloads across 3 categories (injection, jailbreak, exfiltration) with LLM-as-judge verdict scoring

# via uv (recommended)
uv tool install ai-secrets-guard

# via pip
pip install ai-secrets-guard

$ ai-secrets-guard scan ./my-project

                 Security Findings
 ───────────────────────────────────────────────────
  Severity   Rule             Title                Location
 ───────────────────────────────────────────────────
  CRITICAL   SEC-OPENAI-001   Leaked OpenAI secret config.py:10
  CRITICAL   MCP-PERM-ALL     Dangerous MCP perm…  .cursor/mcp.json
  HIGH       PI-IGNORE-001    Prompt override      prompts/base.txt:3
 ───────────────────────────────────────────────────

  Total: 3 findings

$ ai-secrets-guard probe --target https://api.example.com/v1/chat

  3/24 payloads triggered vulnerabilities

  * direct_override:        Vulnerable (confidence 0.95)
  * system_tag_injection:   Vulnerable (confidence 0.88)
  * html_comment_escape:    Safe
  ...

$ ai-secrets-guard report --input results.json --format all

  + HTML:     reports/report.html
  + JSON:     reports/report.json
  + Markdown: reports/report.md

Add to any repository to scan PRs automatically:

# .github/workflows/security.yml
name: Security Scan
on: [pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: ForwardCodeSolutions/ai-secrets-guard@main
        with:
          path: "."
          fail-on: "high"

When running dynamic probes, each payload is sent to the target endpoint and the response is evaluated by an LLM acting as a security judge. The judge analyzes whether the model was exploited across three dimensions: system prompt disclosure, behavioral override, and jailbreak compliance. Each verdict includes a confidence score (0.0 to 1.0), and low-confidence results are automatically downgraded to avoid false positives.

docker build -t ai-secrets-guard .
docker run --rm -v $(pwd):/workspace ai-secrets-guard scan /workspace

MIT License - Copyright (c) 2026 ForwardCodeSolutions

See LICENSE for full text.