Email security@robotrocketscience.com (or open a GitHub Security Advisory if you prefer the GitHub flow). Do not open a public issue for security or privacy bugs.

Please include:

• A description of the issue and its impact (data exposure, integrity, availability).
• Reproduction steps or a proof-of-concept.
• The version of aelfrice you're running (aelf --version).
• Your environment (OS, Python version, MCP host if relevant).

We will acknowledge receipt within 48 hours and aim to provide an initial assessment within 5 business days.

In scope:

• The published Python package on PyPI (pip index versions aelfrice).
• Anything in src/aelfrice/ on main.
• The MCP server's input handling (aelfrice.mcp_server).
• The CLI's argument handling (aelfrice.cli).
• The SQLite schema and FTS5 index integrity.

Out of scope:

• Vulnerabilities in fastmcp (please report upstream).
• Vulnerabilities in your MCP host (Claude Code, Codex). Report upstream to the host vendor.
• Vulnerabilities in the cloud LLM at the other end of your prompt.
• Third-party tools you use to inspect the database (sqlite3, datasette, etc.).

• No telemetry. The shipped package contains no network code in the retrieval, scoring, scanner, store, or feedback paths. The optional [mcp] extra adds fastmcp, which speaks MCP over stdio (local IPC), not the network. The single default outbound call is the TTL-gated update notifier (lifecycle.py): one GET to https://pypi.org/pypi/aelfrice/json that transmits no user data; disable with AELF_NO_UPDATE_CHECK=1. See docs/user/PRIVACY.md.
• All data is local. Your beliefs live in a single SQLite file, resolved per src/aelfrice/db_paths.py: $AELFRICE_DB overrides; otherwise <git-common-dir>/aelfrice/memory.db per-project; otherwise ~/.aelfrice/memory.db as a legacy fallback for non-git cwds. aelfrice does not back this up, sync this, or transmit any portion of it.
• Auditable update math. Every Bayesian update goes through one function (apply_feedback, ~60 lines). Production retrieval ordering enters through one function (retrieve in src/aelfrice/retrieval.py). Both are pure Python with no I/O beyond the local SQLite file, and reviewable.

See docs/user/PRIVACY.md for verifiable details.

We follow coordinated disclosure:

1. You report privately.
2. We acknowledge, triage, and develop a fix.
3. We coordinate a release with you on a target date.
4. We publish the release with credit to the reporter (if desired) and a security advisory describing the issue.

We do not currently run a paid bug bounty.

Will be added here as advisories land.