This plugin makes automated safety decisions about AI-agent permissions, so we take security bugs seriously.

Please do not file a public issue for a security vulnerability. Instead, use GitHub's private vulnerability reporting:

• go to Report a vulnerability at https://github.com/Warc0s/opencode-permission-reviewer/security/advisories/new

or email the maintainer directly if the report contains sensitive details.

Include, when possible:

• a minimal reproduction (command + permission policy that triggers it),
• the affected version (package.json → version),
• the OpenCode version you run,
• the expected vs. actual safety outcome.

We will acknowledge within a few days and coordinate a fix and disclosure.

This plugin is a defense-in-depth aid, not a complete sandbox. It reviews ask-classified actions with a second model call and fails safe to manual review when anything is uncertain. It does not replace OpenCode's own permission system, your model provider's safety layers, or good OS-level hygiene. A determinedly adversarial agent may still attempt to mislead the reviewer; the deterministic emergency brake and the untrusted-evidence prompt mitigate but cannot fully eliminate that risk.