apitame is a safety layer between agents and real systems, so security reports are taken seriously and handled quickly.

Please report privately through GitHub security advisories. Do not open a public issue.

Include:

• what an agent or attacker can do that the safety model says it should not (for example run a write in read-only mode, bypass confirmation, read a denied table, or see a secret);
• the config, spec or SQL needed to reproduce it;
• the apitame version and Node.js version.

You can expect an acknowledgement within 3 working days and a fix or mitigation plan within 14 days for confirmed issues. Credit is given in the release notes unless you prefer otherwise.

• Bypassing read-only mode, the write allowlist, confirmation or dry run.
• SQL that passes the guard and modifies data, runs side-effect functions, reads denied tables or escapes the read-only transaction.
• Secrets appearing in tool results, errors, logs or previews.
• The HTTP transport accepting requests it should refuse (Host, Origin or token checks).

• Behavior that requires editing apitame.config.yaml: the config file is trusted.
• Data the configured credentials are allowed to read (see "What apitame does not protect against" in docs/safety.md).
• Vulnerabilities in the APIs or databases apitame connects to.