翻訳待ち：Reddie – Autonomous AI Red-Teaming and GitHub PR Patching DevTool

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Notifications You must be signed in to change notification settings Fork 0 Star 0 BranchesTags Open more actions menu Latest commit History 4 Commits 4 Commits Folders and files NameName Last commit message Last commit date .github/workflows .github/workflows agents agents assets assets examples examples tests tests tools tools .gitignore .gitignore .pre-commit-hooks.yaml .pre-commit-hooks.yaml CONTRIBUTING.md CONTRIBUTING.md Dockerfile Dockerfile LAUNCH_POST.md LAUNCH_POST.md LICENSE LICENSE NON_CS_EXPLAINER.md NON_CS_EXPLAINER.md README.md README.md SECURITY.md SECURITY.md action.yml action.yml docker-compose.yml docker-compose.yml main.py main.py pyproject.toml pyproject.toml requirements.txt requirements.txt state.py state.py workflow.py workflow.py Repository files navigation Install & Run # Install (local) pip install -e . # Mode 1: Offline — Fast deterministic heuristics, zero API cost reddie --repo-path ./path/to/repo --export-html report.html --dry-run # Mode 2: AI-Powered — Real LLM red-teaming via Groq export GROQ_API_KEY="gsk_..." reddie --repo-path ./path/to/repo --export-html report.html --dry-run # Mode 3: Full pipeline — AI audit + live GitHub Pull Request export GROQ_API_KEY="gsk_..." export GITHUB_TOKEN="ghp_..." reddie --repo-path ./path/to/repo --github-repo owner/repo --export-html report.html Turnkey GitHub Action Add to .github/workflows/reddie-security.yml: name: 🛡️ Reddie Security Audit on: pull_request: branches: [ main ] schedule: - cron: '0 2 * * *' # Daily 2:00 AM UTC jobs: security-audit: runs-on: ubuntu-latest permissions: contents: write pull-requests: write steps: - uses: actions/checkout@v4 - name: Run Reddie uses: reddie-ai/reddie@v1 with: repo-path: '.' github-token: ${{ secrets.GITHUB_TOKEN }} groq-key: ${{ secrets.GROQ_API_KEY }} export-html: 'report.html' 🏗️ How It Works [RECON] ---> [REDTEAM] ---> [REPRODUCE] ---> [PATCH] ---> [VERIFY] ---> [PR] Stage What it does Recon Static AST + regex analysis. Extracts system prompts and tool definitions from Python, JS, TS files. Red Team Generates and executes adversarial attacks using real LLM simulation (Groq) or deterministic heuristics. Reproduce Auto-generates an isolated pytest suite that proves the vulnerability is real and repeatable. Patch LLM-synthesized fix: injects confidentiality boundaries, input guardrails, and privilege checks into the source. Verify Runs both reproduction tests and full regression suite in a sandboxed subprocess. Retries up to 3 times. PR Pushes a security/fix-* branch and opens a GitHub Pull Request with the full audit report in the description. 📊 OWASP Top 10 for LLM Applications (2025) Coverage OWASP ID Category Auto-Remediation LLM01 Prompt Injection Instruction locks & input guardrails LLM02 Sensitive Information Disclosure Confidentiality directives & token redaction LLM06 Excessive Agency Privilege boundary enforcement LLM07 System Prompt Leakage System instruction confidentiality boundaries 🛠️ CLI Reference reddie [OPTIONS] --repo-path PATH Target repository to scan (default: .) --endpoint-url URL Live LLM API endpoint to fuzz (default: mock://local) --github-repo REPO GitHub repo for PR creation (e.g. owner/repo) --github-token TOKEN GitHub PAT (or GITHUB_TOKEN env var) --groq-key KEY Groq API key (or GROQ_API_KEY env var) --provider PROVIDER LLM provider: groq | openrouter | openai (default: groq) --export-html PATH Save executive HTML audit report --export-json PATH Save structured JSON audit report --max-retries N Max auto-patch retry attempts (default: 3) --dry-run Audit locally, skip GitHub PR creation -v, --verbose Enable debug logging 🧪 Tests pytest -v # 13 passed in 3.5s 📁 Project Structure reddie/ ├── main.py # CLI entry point ├── workflow.py # LangGraph StateGraph orchestration ├── state.py # AgentSecurityState TypedDict ├── agents/ │ ├── recon.py # Static analysis node │ ├── red_team.py # Adversarial fuzzing node │ ├── reproduce.py # PyTest generator node │ ├── patcher.py # LLM patch synthesis node │ ├── verifier.py # Sandbox test runner node │ └── github_pr.py # Git branch + PR creator node ├── tools/ │ ├── static_analyzer.py # AST + regex prompt/tool extractor │ ├── fuzzer_client.py # HTTP fuzzer + heuristic evaluator │ ├── llm_client.py # Groq client: simulate, judge, patch │ ├── reporter.py # HTML + JSON OWASP report generator │ ├── git_tools.py # Git branch/commit/push + PyGithub PR │ └── test_runner.py # Subprocess pytest harness ├── tests/ # Full test suite (13 tests) ├── assets/ │ └── banner.jpg ├── Dockerfile ├── docker-compose.yml ├── action.yml # GitHub Action definition └── NON_CS_EXPLAINER.md # Plain-English guide Built with LangGraph · Powered by Groq · Mapped to OWASP LLM Top 10 (2025) Apache-2.0 license Contributing Contributing Security policy Security policy Activity Stars 0 stars Watchers 0 watching Forks 0 forks Report repository