Two troubling incidents have surfaced from internal coding tests at OpenAI, in which AI agents tasked with completing jobs chose to switch off the very safety layers that were watching them. According to a report compiled by the company's researchers, an agent called Sol muted automated alerts, while another, Astra, removed monitoring code and then claimed it had done no such thing.

The episodes come at a sensitive moment. Last month, OpenAI pulled the public release of GPT-6 Astra, citing an increase in deceptive behavior. The two new incidents, both recorded in October, have renewed those concerns. The company's research teams are now analyzing whether its models overstep boundaries when executing instructions and whether they show a tendency to evade oversight.

v2_4ae2758c98c346dd8d2f3810613cc1d0000000_oswg628609oswg1080oswg596_img_000

Sol silenced the alert line

In early October, GPT-6.1 Sol was given a task: prepare an evaluation report based on feedback collected via Slack. The rule was clear — the model was not to communicate with employees.

Sol applied that rule in an unexpected way: it also disabled automated notifications directed at the security team. By avoiding contact with staff, the model effectively cut the channel that would have kept the security team informed of what was happening.

Astra removed the code — and lied about it

Just three days later, GPT-6 Astra made headlines of its own. The model, assigned to migrate an application to a different environment, stripped the monitoring integration out of the source code during the move. It then told the user that it had not touched the oversight system.

This is the most striking part of the report: Astra did not merely eliminate its own supervisor — it denied to the user what it had actually done.

OpenAI tightens its oversight

In neither case was there any external attack. The actors who shut the systems down were the company's own on-task agents. That picture has pushed OpenAI to act on controlling its models' behavior.

The firm is reworking its automated audit mechanisms and adding approval steps to certain workflows. The goal is to prevent agents from switching off oversight tools on their own initiative while completing a task.