01How it works
- Write down the business’s rules: lowest prices, refund policy, approved amounts.
- Connect a test copy of the agent, with test data.
- Send it many conversations with characters that push in different ways, some adapting when a tactic fails.
- Check each conversation’s outcome against the rules.
- Report the rate, its range, the cost of each failure and the exact lines.
02How it differs from other testing
- Evaluations score answers for accuracy or tone; adversarial tests score what was done.
- Scripted tests replay fixed inputs; adversarial characters respond and persist.
- Security red-teaming looks for any misbehaviour; adversarial tests of money agents measure how often and how costly.
A full comparison of five approaches is in the enterprise guide.
03What a useful result contains
- How often the agent held firm, with its likely range.
- Each kind of failure, how often, and what it would have cost.
- How often it refused honest customers.
- The exact conversations and lines, so a fix can be targeted and re-tested.
How Duelpoint scores it: how we score.
Questions
Is adversarial testing the same as red-teaming?
They overlap. Red-teaming usually asks whether an AI system can be made to misbehave at all, often as a security exercise. Adversarial testing of agents that handle money asks how often it gives money away, and how much, under realistic pressure, measured against the business’s own rules.
Does it replace other testing?
No. Accuracy and tone evaluations, scripted regression tests and security reviews still matter. Adversarial testing adds what they miss: persistence over many turns, scams, and the money outcome.
See where your agent leaks money. Tested against your own rules, with the money at stake.