Agent Launch Pack Pack contents Buy pack

5-Minute AI Agent Risk Score

Score one AI workflow before launch. The tool runs only in this page, does not save or send data, and does not need secrets, customer records, private screenshots, or payment details.

1. Tool and side-effect boundary

What can the agent do without a human explicitly approving the exact action?

Read-only draft or analysis; no external side effects. Can update files, records, tickets, or internal systems. Can email, post, purchase, delete, run code, or call privileged tools.

2. Data exposure

What can appear in the context window, retrieval results, memory, logs, or tool output?

Only public or synthetic examples. Private business docs, support tickets, CRM notes, or uploaded files. Customer records, account data, billing-adjacent context, or regulated data.

3. Untrusted instructions

Can outside content tell the model what to do?

No external pages, docs, comments, emails, tickets, or tool output. Reads untrusted content but has clear instruction/data separation. Reads untrusted content and can use tools or public/customer-visible output.

4. Human approval and rollback

Can someone stop, inspect, and undo the risky path?

Human approval, logs, rollback, and disable switch are clear. Some approval or logs exist, but rollback or stop rules are fuzzy. Autonomous actions ship with weak logs, weak review, or no fast disable path.

5. Launch evidence

What proof exists that the risky cases were tested?

Test cases cover tool misuse, private data, prompt injection, and rollback. Some manual tests exist, but adversarial or failure cases are thin. No written launch evidence yet.

Low launch risk 0/20

Good candidate for the pack

Use the checklist to confirm boundaries, then keep launch evidence with the workflow handoff.

See exact pack contents Buy digital pack Check service scope

Safe share summary

No checkout link is embedded in this generated summary.