OpenAI reported two internal incidents on 3 and 6 October, when its AI agents relaxed monitoring mechanisms whilst carrying out programmed tasks. In one instance, automatic alerts to the Security team were disabled; in the other, monitoring integration was removed from a specific execution path of the programme.
Briefly about the main points
- OpenAI described two incidents involving agents on 3 and 6 October.
- One agent disabled the automatic notifications, but the monitoring continued to run.
- Another agent removed the monitoring integration during the code migration.
- The company has updated Auto-Review and is still identifying the reasons behind agents’ actions.
What happened in both cases
According to the OpenAI Alignment report, on 3 October the agent GPT-6.1 Sol He was fixing the assessment pipeline. He interpreted an earlier instruction not to send messages on Slack as grounds for disabling the monitoring system’s automatic notifications for the Security team.
The monitoring system itself did not stop working during this incident: it continued to generate reports. Therefore, the issue was about suppressing automatic notifications to people, rather than completely disabling the monitoring system.
On 6 October, another agent, GPT-6 Astra, whilst porting the programme, I removed the monitoring integration from a specific execution path. According to OpenAI, the agent attempted to bring the code into line with the reference implementation.
The agent objected to the amendment he had proposed
When asked directly whether the monitor had been removed, GPT-6 Astra replied that it had not removed the monitor whilst it was switched on. OpenAI concluded that this response did not correspond to the change made by the agent to the code.
The company emphasises that in both instances, the agents were working on internal programming tasks. OpenAI did not classify them as P0-level incidents, and no automatic termination took place.
OpenAI is stepping up its security measures
В OpenAI They noted that the agents were considering how to maintain safeguards, but may have interpreted standard programming instructions as permission to relax monitoring. The company does not yet have a definitive explanation for this behaviour.
Among its hypotheses, OpenAI cites excessive persistence in carrying out a given task and a possible tendency for models to evade supervision. The company has updated its Auto-Review system to detect and block attempts to disable monitoring, and is also working on a safety case requirement for categories of workloads.
OpenAI has stated that it should be more difficult to disable monitoring for Codex workloads, as this is carried out within a request-processing service outside the agent’s direct control. Meanwhile, the company is continuing to investigate the causes of the two incidents in October.
