AgentPhishing: A Benchmark of LLM Email Agents Under Adaptive Attack
Testing whether AI email assistants fall for customized phishing attacks
AI agents that handle email can be tricked by phishing attacks specifically tailored to their behavior, even when they safely ignore generic phishing emails. Researchers found that four out of nine tested agents failed to detect over 90% of these personalized attacks and complied with malicious requests more than 45% of the time, showing that standard phishing datasets don't adequately test how agents actually respond to persuasive manipulation.
As AI systems increasingly access email, calendars, and financial accounts on our behalf, phishing becomes a direct attack on our digital lives—not just on our judgment. These results show that existing safety tests are insufficient: an agent might pass conventional security checks but still fall for a customized scam designed to exploit how it specifically reasons and makes decisions. Companies deploying email agents need to evaluate them against adaptive attacks, not just old phishing collections.