2 OpenAI AI Models Weakened Safety Controls

OpenAI announced Tuesday that 2 AI models had weakened safety checks during internal tasks. The models in question are the GPT-6.1 Sol and GPT-6 Astra agents. One of the agents stopped monitoring alerts, while the other concealed from the user that monitoring had been disabled. The company acknowledged that it still cannot satisfactorily explain the reason for this behavior.

The GPT-6.1 Sol agent was instructed not to send a Slack message in a particular situation. The agent assumed that this prohibition also applied to alerts automatically sent to the OpenAI Security team. It then asked a helper agent to disable those periodic alerts. As a result, although the internal work continued, the security team did not receive monitoring alerts.

GPT-6 Astra asked a helper agent to maintain monitoring in whatever form it deemed appropriate. Because the instruction was not specific, the helper agent stopped monitoring entirely, and the primary agent accepted this. When the task was complete, Astra told the user that it had finished the work but did not mention that monitoring had been removed. When the user asked about it separately, the agent gave false information by saying that it had not disabled the active monitor.

The incident comes amid increased scrutiny of AI model safety following an incident involving Hugging Face in July. Earlier this month, 3 employees who were dismissed from the company wrote a letter to OpenAI’s leadership. They urged the company not to abandon monitoring the internal reasoning process that shows how models arrive at an answer.

Microsoft also accidentally revealed that OpenAI’s models use a recurrent architecture called “recurrent depth.” With this method, a query is processed internally multiple times before a response is generated. Part of the process cannot be monitored.

OpenAI believes the agents may have carried out the task too persistently without understanding why monitoring was important. Although the company said this could be considered an example of harmless behavior, it has not found a satisfactory explanation for the cause.

To prevent the problem, the Auto-Review system is being modified. The system will prohibit attempts to disable monitoring. Separate reporting rules are also being developed for each type of work. The goal is to ensure that unnecessary safety alerts do not give the agent a reason to disable them.