jev-guard is a guardrail, not a sandbox. It sends tool calls and tool results to Jev (TypeSafe, or Vercel AI Gateway) and turns the answers into deny / ask / flag decisions. It fails open by default so an outage never freezes your agent; set JEV_GUARD_FAIL_CLOSED=1 if you prefer the opposite.

Report a vulnerability privately through GitHub security advisories rather than a public issue. Bypasses of the guard that come from a host's hook semantics (for example a tool path a host doesn't route through hooks) are worth reporting too; they belong in the README's limitations even when they can't be fixed here.