Quick answerA genuine, curious or skeptical customer probing an AI agent with hypothetical questions is not an adversarial actor and should not be treated like one; the agent should answer honestly, including saying when a scenario is outside what it can meaningfully judge, because a skeptical customer who gets a straight answer usually converts into a more trusting one, while evasiveness confirms exactly the doubt they came in with.
Not the same as the two adversarial cases already covered
This is a genuinely different situation from detecting an undisclosed adversarial audience such as a competitor or journalist and from a customer who explicitly identifies themselves as a competitor or journalist. Both of those posts are about people probing the agent for reasons unrelated to their own use of the product, often to expose a weakness or gather competitive intelligence.
The customer in this post is a real prospective or existing customer. They are testing the agent because they are trying to decide whether to trust it, not trying to catch it out for someone else's benefit. Treating this customer with the same caution appropriate for an adversarial audience is a mismatch that actively damages the sale or the relationship.
Why skepticism at this stage is normal, not a red flag
A customer asking "what would you do if I asked you to do something against policy" or "can you actually access my account or are you just guessing" is doing exactly what a careful buyer should do before trusting an automated system with something that matters to them. This is closer to due diligence than to an attack. An agent that responds evasively, or that gives a scripted-sounding non-answer, confirms the exact worry that prompted the question in the first place: that the system is not actually reliable or transparent about its own limits.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
The right response pattern: honest, specific, and comfortable admitting limits
The agent should answer hypothetical and trick questions the same way it would answer a real request: honestly and specifically. If the honest answer is "I do not have access to that information" or "I am not able to make that judgment call, a human would need to review it," the agent should say exactly that, the same posture covered in designing an agent that can say I don't know without losing trust. A skeptical customer who gets a precise, non-defensive answer, including an honest admission of a limit, typically walks away more confident in the system, not less, because the response demonstrated exactly the reliability they were testing for.
What not to do: treat curiosity as a threat to manage
Avoid building special-case detection logic that tries to identify "testing" behavior and route it differently, the same trap that applies to adversarial-audience detection generally. A genuinely curious customer and a genuinely confused one can phrase questions almost identically, and misrouting a real customer's honest question into a defensive or restricted response path creates exactly the bad experience the whole exercise was meant to avoid. Design one consistent, honest response policy that works whether the question is a real request, a hypothetical, or a trick question, and let that consistency itself be the answer to the customer's underlying test.
FAQ
How can you tell a genuinely curious customer from an adversarial one if the questions look similar?
In most cases you do not need to, and should not try. Answer honestly either way. The few cases where the distinction actually matters, such as active competitive intelligence gathering, are covered separately in the adversarial-audience posts, and even then the agent's own answers should stay consistent.
Should the agent acknowledge that it is being tested?
It can, briefly and without defensiveness, for example by simply answering the hypothetical as asked. Calling out the test explicitly is usually unnecessary and can come across as evasive.
Does this apply to pre-sales prospects as well as existing customers?
Yes, and arguably more so, since a prospect testing the agent's honesty and limits before signing up is making a trust decision that directly affects whether they become a customer at all.