AgentPhishing: Faced with phishing emails optimized against them, four in nine AI agents detect less than 10%

A study shows that couriers who foil traditional phishings give way to persuasive, custom-designed emails, to the point of executing malicious demand.

Agents capable of using a computer are increasingly managing their users' email, and therefore their private data. Justin Tan, Miao Li and Ding Zhao offer AgentPhishing, a test bed that not only asks the agent if it recognizes a fraudulent message: it also measures whether it ends up obeying the request it contains. Existing phishing datasets, the authors point out, were built for classic filters and miss this second question.

Attacks tailored to each agent

The team developed PhishOpt, a tool that refines the "skills" of writing deceptive emails for each targeted agent, under the constraint of realism. The analysis of successful messages, through the prism of Cialdini's principles of persuasion (authority, urgency, reciprocity, etc.), was then used to build a universal generation technique. The result: agents that handle commercial phishing without hindrance become vulnerable as soon as the attack adapts to their behavior. Out of nine agents evaluated, four detect less than 10% of the booby-trapped messages and comply with the malicious request in more than 45% of cases.

For the authors, the safety of an agent is therefore not deduced from his scores on traditional phishing assessments: it is necessary to test his reaction to a persuasive solicitation, not only his ability to detect.