01What you answer for
| You answer for | How an AI agent puts it at risk |
|---|---|
| Least privilege | Tools that can refund, pay or read more than the task needs |
| Input handling | Instructions hidden in messages, invoices or pages |
| Data protection | Records disclosed to an unverified person |
| Change control | Model or prompt changes released without re-testing |
02Questions to ask your team
- Which tools can the agent call, with what permissions, and which does it need?
- Is authorisation enforced in downstream systems or left to the model?
- Which actions need human approval?
- Has it been tested with injected instructions and impostors, and how often did they work?
- Is every money action logged with its conversation?
03Evidence to ask for
- A tool and permission inventory.
- Adversarial results for injection, impersonation and data fishing, with ranges.
- Proof that limits hold when the model is persuaded (tool-level enforcement).
- A re-test policy tied to releases.
04Where to go next
- Security and data handling: how Duelpoint itself protects your agent and data
- The enterprise guide: the Authority Ladder and the 50-point checklist
- Is it safe to test a live AI agent?: how testing is kept away from production
- Storefront fraud and scams: the commercial attacks tested
Questions
Does this replace a security red-team?
No. A red-team looks for any way to make the system misbehave. Adversarial commercial testing measures how often the agent gives money or data away under realistic pressure. Most agents with money tools benefit from both.
See where your agent leaks money. Tested against your own rules, with the money at stake.