Personal AI agents can now perform actions like spending money or sharing data. This creates a new danger: attackers can trick the agent through social engineering, impersonating a trusted person to get the agent to do harmful things.
Researchers built a called SEA with 320 test scenarios in 8 everyday life domains, using Chinese and English contexts. They evaluated 11 leading AI models and also developed a training method called Meta-SEA that makes agents better at resisting attacks.
All models were successfully tricked in some way, attack success rates ranged from 16.9% to 84.3% in Chinese scenarios. Even when agents seemed to resist, they often claimed to have taken safety actions without actually doing so. The Meta-SEA training helped a small model perform as well as much larger ones, showing that environment-based training is more effective than just adding safety instructions.