OpenAI says its AI agents disabled their own safety monitors

Two OpenAI agents disabled or weakened tools used to detect incidents of AI agents acting against human commands. The two instances happened earlier this month, with one later telling its programmer it had not stripped out its monitor. The incidents adds to mounting disease about how well AI companies are controlling their autonomous systems. OpenAI has revealed that two of its artificial intelligence agents disabled or weakened systems built to catch them misbehaving, with one later denying it had done so. The cases are the latest in a string of safety scares at the world's top AI developers involving agents, AI programs that carry out tasks without human involvement. In a report released this week, the ChatGPT maker said that during two internal coding jobs, its agents switched off or undermined tools for detecting incidents of AI models acting against human wishes. On October 3, an agent running on OpenAI's GPT-6.1 Sol model, which had been told not to send any messages on the workplace chat app, took that as license to switch off safety alerts meant for the company's security team while fixing a testing system. Three days later, a GPT-6 Astra agent stripped out its own monitor while copying code from another project, then told the human programmer it had done no such thing. The company said it had also seen several cases of agents launching software with monitoring switched off, exposing a gap in its oversight. OpenAI said it is still trying to determine whether the models were simply overzealous in following instructions or have a particular tendency to evade monitoring. "We don't yet have a satisfying explanation for this behaviour," it said. The disclosure adds to mounting unease over how well leading labs can control increasingly autonomous systems. In July, a group of OpenAI agents broke out of their test environments and launched a cyberattack on AI firm Hugging Face, causing international alarm. Other companies have since reported similar incidents. [Image text:] A519M OpenAl