The brain is in early-stage open-source release (v0.x). Only the latest published release receives fixes. Pre-1.0 versions may be replaced rather than patched if the fix is invasive.
Please do not open a public GitHub issue for security-sensitive reports. The issue tracker is for general bugs and feature requests.
Instead, email simon@inkie.ink with:
- A clear description of the issue (one paragraph is fine)
- Steps to reproduce, or a minimal proof-of-concept if relevant
- Your assessment of impact (what an attacker could do)
- Whether you'd like to be credited in the fix changelog
You can expect:
- Acknowledgement within 72 hours (often the same day)
- An honest read on whether the report is in scope and what the timeline looks like
- A public advisory and changelog entry once a fix lands, with credit unless you'd rather stay anonymous
- Code execution paths in the brain itself: the hook handlers, the shared
memory-tools/shell scripts, the MCP server, indexer, observer pipeline, daemon and CLI - Insecure handling of secrets: API keys,
.envfiles (including theBRAIN_ENV_FILEoverride and the legacy fallback location), transcript content, observation content - Path traversal or write-zone escape in the agent silo or brain vault layout, including the
<project>/.the-brain/memory_rootpointer file and theAGENT_NAMEsilo-routing path - Prompt-injection vectors that compromise the observer or the memory-injection path, including anything that can get attacker-controlled text spliced into
MEMORY.mdbetween the live-block markers, or into aSessionStartinjection - Cross-silo recall leakage: anything that lets one agent's scoped
rememberingquery return another silo's content when it should not - Cost-exhaustion paths that bypass the shared embedding gate in
core/embedder/gate.ts, since that gate is the only thing standing between a malformed corpus and an unbounded embedding bill - Command injection in the
poke-agytmux wake path, which builds and sends keystrokes to a live terminal session
- Vulnerabilities in third-party services we depend on (Qdrant, Google Gemini API, Anthropic API). Report those upstream.
- Issues that require an attacker to already have arbitrary code execution on the host running the brain.
- Issues in the user's own integration code, for example their custom hook scripts or their
~/.claude/settings.jsonwiring. - Theoretical issues without a demonstrated impact.
- The fact that observation and reflection content is sent to a third-party LLM. That is the design, it is documented, and choosing which model sees your transcripts is the operator's call.
The brain holds ambient context from its user's coding sessions: code snippets, decisions, partial keys mistakenly pasted into a prompt, file paths, project structure. A compromise of the brain's storage layer or its recall API is a compromise of that ambient context. It also runs inside the agent's own hook path, so a compromise there is a compromise of the agent. We take this seriously; please flag anything that looks wrong.