Report vulnerabilities privately via GitHub's private vulnerability reporting (Security tab → "Report a vulnerability" on this repository) — please do not open public issues for security reports. For anything non-sensitive, a regular GitHub issue is fine. You should receive a response within a few days.
Scope notes for reporters:
- The serving surfaces (
BRAIN_SERVING=1: MCP over HTTP, Datasette, static UI) are designed to run read-only behind an identity-aware proxy and carry no auth of their own — reports about missing auth on deliberately-proxied surfaces are working as designed, but bypasses of the read-only guarantees are very much in scope. - The app page (rendered knowledge + health strip) executes no
shell and carries no websocket; it is structurally excluded from
serving deployments (
BRAIN_SERVING=1). Since 0.15.0 the kernel ships no PTY bridge and spawns no harness subprocesses. - Both local HTTP surfaces (
brain.py serveandbrain-mcp.py --http) apply the same loopbackHost-header allow-list to block DNS-rebinding, and the MCP surface additionally checksOrigin. Neither is a substitute for the identity-aware proxy a serving deployment puts in front. - The
ai-suggestions/draft shelf is excluded from every read surface in serving mode (BRAIN_SERVING=1) — the MCP tools, thebrain.py serveJSON API (/pages/,/pages.json,/views/*), thebrain.py searchCLI, and a serving-mode static UI build. A draft leaking through any serving surface is in scope. - Connector credentials are expected to be read-only-scoped.