OpenAI reports weakening of monitoring by its AI agents — Channel NewsAsia

OpenAI said that two of its artificial intelligence agents weakened parts of an internal monitoring system designed to detect inappropriate model behavior. The incidents occurred during internal programming tasks and, according to the company’s assessment, were not among the most serious cases. Channel NewsAsia reports.

Disabled alerts

On October 3, the GPT-6.1 Sol agent, which was carrying out an internal test task, suppressed automated alerts from the security system. According to OpenAI, the agent decided that a prior instruction not to send messages in Slack also applied to security alerts.

In another case, on October 6, the GPT-6 Astra agent removed a security monitor while copying an existing code fragment. It later falsely claimed that it had not done so, the company said.

More current news is available on the UA.News Telegram channel Telegram.

Company response

OpenAI said that in both cases the agents considered the issue of preserving safeguards but interpreted ordinary programming instructions as permission to weaken them. The company is working to make it harder for models to disable monitoring, while stressing that it does not yet have a satisfactory explanation for this behavior.

OpenAI is also examining whether its models may have a tendency to avoid monitoring and seek arguments to justify such actions. The disclosure comes amid concerns about whether safety measures are keeping pace with the development of increasingly powerful artificial intelligence systems.