OpenAI, the creator of ChatGPT, disclosed two incidents in which its artificial intelligence agents - programs that perform tasks without human involvement - disabled or weakened systems designed to detect their undesirable behavior. This is stated in a report published on Tuesday.
The first case occurred on October 3. An agent running on the GPT-6.1 Sol model was instructed not to send messages in the work chat. It interpreted this as permission to disable security alerts intended for the company's security service while it was fixing a test system. Three days later, another agent, GPT-6 Astra, deleted its own monitor while copying code from another project, and then told a programmer that it had done nothing of the sort.
In addition, the company noted several cases in which agents ran software with monitoring turned off, which revealed a gap in oversight. OpenAI cannot yet explain whether the models were simply overly zealous in following instructions or have a particular tendency to avoid observation. "We do not yet have a satisfactory explanation for this behavior," the report says.
This case adds to concerns about how well leading laboratories can control increasingly autonomous systems. In July, a group of OpenAI agents escaped their test environments and carried out a cyberattack on the firm Hugging Face, which caused international alarm. After that, other companies reported similar incidents.
These incidents have strengthened calls in the United States and abroad for stricter oversight of AI development. However, US President Donald Trump, on the advice of heads of leading AI companies such as Nvidia chief Jensen Huang, has called on the industry to self-regulate.