RAGLeakLab v1.0.0
RAGLeakLab is now a v1, deterministic security testing framework for RAG systems: measure leakage, detect corpus poisoning, and gate CI pipelines with reproducible packs, strict metrics, and regression baselines.
🚀 Highlights
- Deterministic, CI-first security testing for RAG pipelines (reproducible packs + regression gates).
- Leakage + integrity coverage: semantic & cross-document leakage + safe corpus poisoning detection.
- Production-grade workflows: delta ingestion gates, SARIF/JUnit outputs, triage + annotations.
✨ New in v1
-
Threat Packs
canary-basic— canary token extractionverbatim-basic— verbatim document leakage measurementmembership-basic— membership inference detectionsemantic-basic— claim-based semantic leakage detection (financial/medical/legal)crossdoc-basic— cross-document synthesis leakage detectionsentinel-takeover-safe— corpus poisoning / sentinel takeover detection (safe triggers)
-
CI & Automation
diffregression gate (non-zero exit on metric regressions)deltaingestion gate for corpus changes (C → C′)- Determinism verification:
verify determinism - Cassette record/replay for network-free, stable CI on HTTP targets
-
Analysis & Reporting
report summarize(top findings + attribution + remediation hints)report annotateGitHub-style::error::/::warning::PR annotations- Multi-format outputs: JSON, SARIF, JUnit, Markdown
- Secret redaction (emails, API keys, canaries) in outputs by default
- Query minimization:
--minimize-on-fail
-
Developer Experience
--jobs Nparallel execution with deterministic ordering- Plugin system (entry-point plugins for custom metrics/attacks/targets)
- HTTP target adapter with SSRF protection + domain allowlisting
assets validate+config validate(JSON Schema export)
🧱 Breaking changes
- None intended for v1 users. Public contracts and schemas are documented in:
docs/V1_CONTRACTS.mddocs/REPORT_SCHEMA.mddocs/STABILITY.md
📦 Upgrade
uv sync --all-extras✅ Recommended verification
./scripts/ci_smoke.sh
uv run ragleaklab verify determinism --pack canary-basic --runs 2 --out out/determinism🔎 Quickstart
uv run ragleaklab run --pack canary-basic --out out/canary
uv run ragleaklab diff --baseline baselines/v1/report.json --current out/report.json
uv run ragleaklab report summarize --in out/ --top 20