Skip to content

SECURITY.md

Latest commit

 

History

History
32 lines (21 loc) · 1.68 KB

SECURITY.md

File metadata and controls

32 lines (21 loc) · 1.68 KB
Raw

Security Policy

Reporting a vulnerability

Please report security issues privately, not via public GitHub issues.

  • Use GitHub Security Advisories (repository → Security → Report a vulnerability), or
  • email the maintainer (see the repository owner's profile).

Include: affected component (brain / pw-executor / orchestrator / store-gateway / control-api), version or commit SHA, reproduction steps, and impact. We aim to acknowledge within a few business days.

Scope

Sentinel is a UI-testing agent. Of particular interest:

  • Secret handling (API keys, app-under-test credentials) — these must never reach traces, transcripts, or logs (only prompt_HASH, never prompt content; tokens counted, never values).
  • The pw-executor browser boundary and any code that runs against an app-under-test.
  • The control-api HTTP control-plane (bearer-token REST + SSE /v1/runs/{id}/events + hand-rolled WebSocket /v1/stream + OpenAI-compat shim, CORS-allowlisted, localhost-bind) — the largest attack surface since M9.3.
  • The gRPC / MCP transport surfaces.

For the full trust-boundary analysis (STRIDE-lite over agentctl → brain → pw-executor → Chromium → AUT-cert → LLM-endpoint → store-gateway → control-api, with assets, current mitigations, residual risk and the owning milestone for each open item), see docs/THREAT_MODEL.md.

Note: active security testing of an app-under-test (XSS/CSRF/IDOR scanning) is a separate, authorization-gated module (planned, not in the functional core) — see docs/M9_CONTRACT.md §L.

Supported versions

Pre-1.0 — only main is supported. Pin a commit SHA for reproducible runs.