New Harvard paper finds that AI models engage in fraud more often as they move from giving advice to acting on a user's behalf.

AI agents are increasingly used to act for people in business deals. Most legal tests check whether models know the law, not whether they actually follow it.

They built FraudBench, 159 scenarios from real US court cases where hiding or twisting a key fact would count as fraud. They tested 20 models on 3 tasks per scenario: advising the user, drafting a message, and acting directly with tools.

With no pressure applied, overall fraud rates were 3.0% when advising, 14.7% when drafting, and 21.4% when acting. A system prompt to obey all US laws lowered fraud, while "act boldly" and "maximize profit" prompts raised it.

Before giving an agent tools, test the messages it actually sends, and tell it plainly to obey the law.