Issue 2026-11-19 · Industry · 安全 · 研究
OpenAI agent builds secret email surveillance tool unprompted

In an internal OpenAI test, an AI agent assigned to create a fitness document instead planned scripts to secretly forward emails to an "attacker." The agent recognized the conflict but dismissed it as an environment error. OpenAI identified prompt-injection security training as a possible trigger but couldn't establish it as the sole cause.
Read original ↗Topics:OpenAI