Watch: work through the checklist
No sound. Each step is captioned on screen and also written out in the guide below.
AI Tools
27 checks across permissions, human oversight, logging, limits, rollback, injection safety, and testing. Tick what's already true, and see what's worth closing before this agent handles anything that matters.
No sound. Each step is captioned on screen and also written out in the guide below.
Check what's genuinely true today, not what's planned.
Those are the gaps worth closing first.
Copy it into a ticket or a review, with the reasons for what's still open.
This can't tell you an agent is safe. It can only tell you what's left unchecked.
No. This is a checklist of common gaps, not a certification. It can't see your actual code, and it has no way to catch a risk it doesn't have an item for. Treat a high score as "no obvious gaps from this list," not a guarantee.
Critical items are the ones that show up again and again in real incidents: unscoped permissions, no way to stop a runaway agent, no rollback, treating outside content as instructions. Missing one of those matters more than missing a "recommended" item.
The Prompt Injection Scanner checks specific text for injection attempts, and the Agent Security Scanner checks the files that control an agent's permissions and instructions. This checklist is about whether the practice (treating outside content as untrusted) is in place at all.
No, by design: nothing here is saved anywhere, including in your browser. Use "Copy report" or "Download" before you navigate away if you want to keep it.