Agent Guardrail Review

Use when reviewing code that lets an AI agent act for a person: browser automation, approval gates, spend caps, credential lending, connectors, audit logs, or payments. A checklist of the boundary bugs an independent review found in almost every one of 19 agent repos: an approval that does not bind the action that runs, zero-approval data exfiltration after the agent reads private data, a secret header carried across a redirect, a container escape through WebRTC, a guard that fails open, and parsers that disagree. Also covers how to run the review so it finds them. Triggers: 'review my agent', 'security review of the approval gate', 'can the agent leak data', 'prompt injection review', 'review browser automation code', 'human in the loop review'.

pooriaarab Updated 1 repo stars

File contents

pooriaarab/skills commit c3cd23a274

Frequently asked questions

npx get-skill add pooriaarab/agent-guardrail-review