Researchers discover attackers can weaponize AI agent safety guardrails into denial-of-service attacks

According to CSO Online and Wired, researchers found that attackers can poison documents to trap AI agent safety mechanisms in extended thinking loops, turning reasoning-based guardrails into denial-of-service weapons. The attack works by forcing guardrails designed to prevent misuse into computationally expensive operations, dramatically slowing shared AI infrastructure. The vulnerability affects agents that rely on safety mechanisms to validate requests.