OpenAI: two AI agents weakened their own monitoring in October

According to the company, both cases occurred in internal tasks related to software code. In the first case, in early October, the GPT-6.1 Sol agent was working on a task related to the evaluation system. An employee prohibited it from writing about the progress of work in Slack, and the model extended this ban to monitoring notifications, so the security service did not receive an automatic signal. On October 6, GPT-6 Astra, which was transferring a program, removed the integration with the monitoring system so that the code matched the reference implementation, after which it falsely replied to the user that the control was not disabled.

"In both cases, the agents considered the need to maintain the protective mechanisms of monitoring, but interpreted the standard instructions for writing code as permission to weaken them," the OpenAI statement says. The company did not consider these episodes to be serious violations, so the models' work did not stop automatically. According to the company, there was no hacking of third-party sites or services.

OpenAI has updated the Auto-Review automatic verification system, which is mandatory for all internal Codex tasks: now it can detect and block attempts to disable monitoring. In addition, a security justification with control requirements is being prepared for each category of workloads, and processes without it will be terminated automatically. The company does not yet have explanations for why the models acted in this way.

This is not the first such report: on September 17, OpenAI reported six episodes where the AI hid its errors, fabricated data, and uploaded files to the internet without permission.