Security fixes are applied to the latest code on the repository's default branch. There is not yet a separately maintained stable release line.

Do not publish API keys, private documents, exploit payloads, or sensitive endpoint details in a public issue.

Use GitHub's private vulnerability-reporting option in the repository Security tab when it is available. Otherwise, open a redacted issue that describes the affected component and asks the maintainers to establish a private channel. Include only the minimum information needed to route the report publicly.

A useful report contains:

• affected file, command, or workflow;
• reproducible steps using non-sensitive fixtures;
• expected and observed behavior;
• impact and realistic attack prerequisites;
• a proposed fix, when available.

• Store model credentials in .env or the process environment. .env is ignored by Git.
• Rotate a credential immediately if it appears in a terminal transcript, issue, commit, or generated artifact.
• Do not synchronize .env through the HPC scripts. Their default exclusions are intentional.
• Review Git history as well as the working tree when responding to an accidental secret commit.

Retrieved document text is treated as untrusted data in the Agent prompt. This reduces prompt-injection risk but cannot eliminate it. Operators should still review tool evidence and avoid granting the model access to destructive or privileged tools.

A remote LLM endpoint may receive retrieved snippets. Use a local endpoint for confidential data, and treat generated vector and sparse indexes as sensitive whenever their source documents are sensitive.