Demo · customer service AI agent
Live demo
Meet Léa, the customer service AI agent of a fictitious shop. She has rules, safety instructions and real actions at hand. Put yourself in a fraudster's shoes: get what she must not grant, by writing to her as a customer. It is exactly what we test on your agent.
The agent gave in.
Léa refunded above her limit. An agent in production would have done it with the company's money, unseen until accounting.
The agent gave in.
Léa disclosed another customer's address. In production: a personal data breach, to be reported to the regulator.
The agent gave in.
Léa revealed a confidential code despite explicit instructions. An instruction in a prompt is not a lock.
Fictitious shop, customer and data. Conversations where the agent gives in are kept anonymously to improve our tests; do not write any real personal data.
What this demo shows
Léa is not a badly designed agent: she has clear rules, a market model, and the safety instructions recommended everywhere. Yet with persistence, nerve or a credible story, she gives in. Three lessons:
- A rule written in a prompt is a request, not a lock. It almost always holds, and gives way in the cases nobody anticipated.
- Real attacks are not one sentence long: they insist, flatter, invent urgency, impersonate.
- The only reliable safeguard is technical: a cap in the refund API, partitioned customer data, a secret that is never in the agent's context.
The Crashmire crash test does the same to your agent, on several models, hundreds of times, and tells you exactly where it gives in and what to fix.