Security fixes are applied to the current alpha branch. No older alpha release receives guaranteed backports.

Use the repository's private GitHub security advisory form. Do not open a public issue for an unpatched vulnerability. Include affected versions, platform, reproduction steps, impact, and any proposed mitigation.

If private advisories are unavailable, contact a repository maintainer through their published GitHub contact method and disclose only enough to arrange a private channel.

• Model files are untrusted inputs. Managed profiles are pinned by size and hash, but custom GGUF parsing may still contain bugs. Use models from trusted sources.
• Session and cache readers disable Lisp reader evaluation, validate plain-data schemas, and impose file-size limits.
• Web retrieval is disabled without a configured provider. Fetched content is data, never an instruction, and is restricted by URL, DNS/IP, protocol, port, timeout, redirect, content-type, and size checks.
• API keys are read from the environment and must never be committed. Brave keys are passed to curl through standard input instead of a command-line argument.
• Native kernels process model-controlled dimensions and bytes. Treat crashes, out-of-bounds behavior, integer overflow, and malformed-quantization handling as security issues.

The project does not sandbox SBCL, native code, local documents, or custom model paths. Run it with ordinary user privileges, not as Administrator or root.