Only the latest release on PyPI (laya-guardrail) receives fixes.

Please do not open a public issue for vulnerabilities in laya-guard itself. Report privately via GitHub: Security → Report a vulnerability on scs0209/laya-guard.

Expect an acknowledgement within 7 days. Please include a minimal reproduction (input, command, version, and the verdict you got).

• Raw tool_call text, or anything derived from it beyond the documented truncated SHA-256, written to the telemetry log
• The server or a hook adapter executing, or enabling execution of, the content it is judging
• A crafted tool_call that crashes or hangs the server, or that takes the offline fallback longer than its timeout
• An adapter returning allow for a block verdict, or mishandling its agent's hook contract
• Malicious policy.yaml handling beyond the documented trust model (the policy file is trusted local configuration)

• A single input judged wrongly (missed attack or false positive). These are quality bugs — please open an issue with the input and the verdict so it can become an eval case.
• With the server unreachable and LAYA_GUARD_FAIL=open set, calls that are neither regex-flagged nor safe commands pass unjudged — documented and intentional. (By default they are asked about, or blocked where the agent can't ask; a call that passes anyway in that mode is in scope.)
• Exposing the server beyond localhost with --host 0.0.0.0 — it has no authentication by design.