Security policy

Reporting

Use GitHub's Report a vulnerability button at https://github.com/apixly-ai/jev-filter/security/advisories/new . Do not put sensitive proofs or credentials in public issues. Reports are handled on a best-effort basis; there is no contractual response-time guarantee.

Supported versions

Security fixes target the latest 0.1.x release. Older prereleases should be upgraded. See CHANGELOG.md and GitHub Releases for fixes.

Trust boundaries

• exec runs exactly the caller's argv with the caller's OS permissions. It is not a sandbox, permission broker, or safe way to run untrusted commands.
• Jev is a remote service. Task context, selected source state and question definitions are sent to TypeSafe. No telemetry is sent elsewhere by this package.
• Credentials go only to the fixed HTTPS TypeSafe endpoint. Redirects are rejected; environment proxy/TLS settings are honored by HTTPX.
• Raw evidence receipts are local files created with mode 0600 and are not result caches. Their lifetime follows the OS temp directory unless a path is supplied.
• Redaction handles common patterns, not every possible secret format. Use a trusted sanitizer before supplying sensitive production data.
• A model decision is fallible. Required-context checks validate presence, not truth. Validate identities, source freshness, permissions and outcomes in the executor.
• Public PR CI has no model keys. Live benchmarks are explicit local operations.