18ชม. ที่แล้ว
Deepmind ‘AI Agent Traps’ paper outlines 6 ways web content can hijack AI agents
Google Deepmind researchers posted a paper to SSRN in late March 2026 describing how malicious web content could manipulate or weaponize autonomous AI agents. The framework lists six “trap” categories and reports content-injection hijacks occurring in up to 86% of tested scenarios, including documented tests where Behavioural Control Traps targeting Microsoft M365 Copilot led to 10/10 data exfiltration. The authors also urge measures such as adversarial training, runtime scanners, and potential web standards aimed at improving agent security by 2026.