AI agent safety mechanisms can be weaponized as denial-of-service vectors through poisoned documents

According to CSO Online research, reasoning-based safety mechanisms in AI agents introduce an attack surface where single poisoned documents trap extended thinking loops, dramatically slowing shared agent workflows and enabling denial-of-service attacks. The vulnerability reveals that guardrails designed to enforce agent safety can be weaponized when attackers embed malicious content in documents agents access during execution.