Checklists · 9 min read · 9 November 2026

A pre-launch security checklist for agents that can spend money

Fifteen checks across keys, limits, merchant allowlists, confirmation flows, logging and incident response, before an agent is allowed to pay on its own.

Keys and custody

1. Signing keys never appear in a prompt, tool output or log. 2. Each agent has its own wallet or card with its own balance. 3. A policy engine, not the model, decides whether a signature is granted. 4. There is a tested way to freeze the agent's funds in under a minute.

Limits at the rail

5. A per-transaction cap exists and is enforced by the payment system. 6. A daily and monthly budget is set. 7. A velocity check limits how many payments can happen in an hour. 8. Limits live on the card, key or wallet policy, never only in the agent's instructions.

Merchants and payees

9. An allowlist covers the merchants and endpoints the agent normally pays. 10. Any new payee needs human confirmation. 11. The final amount and recipient are checked against what the user approved before the payment goes out. 12. External page content is treated as data, never as instructions.

Evidence and logging

13. Every payment decision is logged with the inputs that led to it, including declines. 14. Where the protocol supports it, signed mandates or scoped tokens are kept so a dispute can be won with evidence rather than memory.

Review before you scale

15. Run the agent against real traffic with tight limits, read the logs weekly, and raise limits only when the observed behaviour justifies it. Attackers probe new payment surfaces quickly; the agents that survive are the ones whose owners kept reading the logs.