Agents18h ago
New Benchmark Tests AI Agent Humanization
A new dataset of mobile touch dynamics reveals that LMM-based agents are easily detected due to unnatural kinematics. Researchers introduced the Agent Humanization Benchmark to model the struggle between detectors and autonomous GUI agents. This framework treats anti-detection as a MinMax optimization problem. Developers must now refine agent movement to avoid platform-level adversarial countermeasures.