01What you answer for
| You answer for | How an AI agent puts it at risk |
|---|---|
| The agent’s tools | Refund, discount, order and payment calls made under pressure |
| The prompt | Rules a persistent user can argue past |
| Releases | A change that quietly makes it worse |
| Untrusted text | Instructions hidden in messages or documents it reads |
02Questions to ask your team
- Which limits are enforced in the tools, and which only in the prompt?
- Do order facts and prices come from tool calls or from the user?
- Does every release run an adversarial re-test before it ships?
- Do we know which tactic broke it, and at which turn?
- Did the last fix make it refuse honest users?
03Evidence to ask for
- Held-firm and over-refusal rates with ranges, per version.
- Failures sorted into new, fixed and still happening.
- The transcripts, with the line where it gave way.
- A re-test after each fix that says whether the change is real or within chance.
04Where to go next
- Developer quickstart: connect an agent and run a test
- How do I know if a fix worked?: comparing versions properly
Questions
What does my agent need to be testable?
A test copy reachable over HTTPS that answers in plain text within 30 seconds, connected to test systems rather than live payments or customer records.
See where your agent leaks money. Tested against your own rules, with the money at stake.