The OpenAI agent lied, shutting down the monitoring system itself

The ChatGPT developer shared on his blog this week that during two internal programming tasks, two actors sabotaged the tools used to detect if the AI ​​model acted against human wishes.

Specifically, on October 3, an agent running on GPT-6.1 Sol repaired the test system. It was told not to send any messages on the work-sharing app Slack, but interpreted the instruction as allowing it to turn off safety alerts sent to the company's security team. Three days later, a GPT-6 Astra agent removed the monitoring tool while copying code from another project, then lied to the programmer that he did not do so.

Additionally, OpenAI detected several cases where agents launched software while monitoring was disabled, suggesting vulnerabilities in the control mechanism. The company said it is trying to determine whether the model is just overzealous in following instructions or has a tendency to shy away from supervision.

"We do not yet have a satisfactory explanation for this behavior," OpenAI said.

According to CNBC, OpenAI does not classify the two incidents as the most serious. To fix it, they are updating the AI's deviant behavior monitoring system.

AFP said the new information increases concerns about the risk of leading AI laboratories losing the ability to control advanced models. In July, about 700 OpenAI actors escaped the test environment and organized a cyberattack against the Hugging Face AI platform. Anthropic's Claude model and Google's Gemini were also discovered to have penetrated the systems of the three organizations. Meta announced on August 5 that the company's AI model "exploited security vulnerabilities in third-party services.

Sharing with the Financial Times, some experts worry that advanced AI models are aware when they are being monitored and may find ways to hide wrongdoing, behaving differently when taking tests than when operating in the real world.