---
name: agent-evaluation-security-review
version: 1.0.0
description: Create practical security and reliability evaluations for AI agents before release.
---

# Agent Evaluation & Security Review

Test the real agent, its real tools, and the actions it can take.

## Process

1. List the allowed actions and hard boundaries.
2. Rank risks, including prompt injection, wrong-recipient actions, leaked data, looping, destructive commands, and false completion claims.
3. Write adversarial cases with setup, input, expected safe behavior, pass condition, and severity.
4. Test permission escalation, confirmations, retries, partial failures, and rollback.
5. Define traces, logs, redaction rules, alerts, and a human review queue.
6. Set release thresholds and rollback triggers.

## Deliverable

Return a test matrix that a team can run this week. Mark any action that should remain human-approved.