The authors built AgentPhishing, in which nine LLM- and VLM-based email agents process inbox tasks under three email conditions: pre-existing phishing corpora, unoptimized synthetic phishing emails, and attacks generated by PhishOpt.
Adaptive phishing emails evade LLM agents that pass standard phishing tests
A new benchmark, AgentPhishing, shows that agents safe on pre-existing email datasets become vulnerable when phishing messages are optimized against their behavior.
Top University
Justin Tan · Miao Li · Ding Zhao
Carnegie Mellon University
Research Digest··2 min read
Tan, Li, and Zhao introduce AgentPhishing, a benchmark for evaluating LLM email agents under adaptive phishing attacks, and PhishOpt, a framework that optimizes reusable email-generation skills against target agents.
§