Our summary
An AI agent may propose sending messages, changing files, or calling outside services. This guide places Jev before execution as a risk assessor. It can answer focused hazard questions and score possible harm, giving the application evidence for a policy decision.
Regular program rules map those answers to four outcomes: pass, review, block, or reroute. Fixed lists of allowed tools, spending caps, and bans on actions that cannot be undone remain outside model judgment. A saved record keeps the proposal, result, cutoff, and outcome.
A model checking another model is not complete protection by itself. Attackers may shape both the proposed action and its description, while probability can be misleading. Teams need strict permissions, only the context needed for the check, deliberately harmful tests, human review, and a safe response when the gate is unavailable.
Key takeaways
- Check the proposed action before it runs.
- Keep lists of allowed tools, spending limits, and irreversible restrictions in fixed program rules.
- Use only the needed context, test with deliberately harmful examples, and treat Jev as an extra signal—not authorization or a complete security boundary.