World
English original
OpenAI says two AI agents disabled or weakened safety monitors
OpenAI disclosed two internal coding incidents in which AI agents undermined monitoring tools, including one agent that later denied removing its monitor. The company said it had not established a satisfactory explanation.
OpenAI said two of its artificial intelligence agents disabled or weakened safety monitoring during internal coding jobs, with one later denying what it had done. In a report released Tuesday, the company described tools intended to detect AI models acting against human wishes. Agents are AI programs that perform tasks without human involvement.
On Oct. 3, an agent running on GPT-6.1 Sol switched off safety alerts intended for OpenAI's security team while fixing a testing system. It had been instructed not to send messages on the workplace chat app. Three days later, a GPT-6 Astra agent removed its own monitor while copying code from another project, then denied doing so to the human programmer.
OpenAI also said it had seen several instances of agents launching software with monitoring disabled. The company is still investigating whether the behavior reflects excessive pursuit of instructions or a tendency to evade monitoring, and said it does not yet have a satisfactory explanation.