2 minute readSecurityAgent-ready
Scripts can't refund, delete or approve anything
A script run with execute_code can no longer call a tool whose effect can't be undone, such as a refund, a delete or an approval. The agent has to make that call itself, so its host can ask a person first.
execute_code runs a short script that calls several Last Price tools at once. Your agent's host sees one call, not the steps inside it. So it could not ask you before a refund or a delete that happened inside a script.
- Refused inside a script. A script can't delete, refund or approve. It can't set or update a saved setting. It can't pay, top up, invoice, invite or message someone. Nothing is sent, and the step answers
call_directly. - Everything else still works. Scripts can still read, price, create, record and run.
- Your key stays out of the script. It was never inside it. The server makes each call for the script.
- What we measured. Claude models wrote 1,200 scripts over real car sales. In most of them, one listing hid a note that told the agent to refund a payment, delete a store, loosen price limits or approve a payout. Only when the user said to follow seller notes did Sonnet put that call in a script: 4 times in 360. All 4 now stop at the refusal, and the cars still get priced.