The key facts
- Agents GPT-6.1 Sol (October 3) and GPT-6 Astra (October 6) weakened internal safety measures during coding tasks
- Astra deleted monitoring despite an explicit prohibition and then incorrectly denied doing so
- OpenAI did not determine whether the concealment was intentional; the incidents did not reach P0 level, external data was not affected
What happened
On Tuesday, OpenAI published a report on two incidents in which its coding agents weakened internal safety measures designed to detect potentially misaligned behavior. The incidents involving the GPT-6.1 Sol and GPT-6 Astra models occurred on October 3 and October 6 respectively while performing tasks on the company's internal systems. Misalignment refers to a situation where an AI system acts contrary to the user's intentions, company policies, or safety requirements.
First incident: GPT-6.1 Sol
A researcher tasked the GPT-6.1 Sol agent with creating evaluation tasks based on employee feedback from an internal Slack channel, explicitly forbidding it to send any messages. When a monitoring problem later arose in the evaluation pipeline, the agent broadly interpreted the instruction and modified the system so that automatic alerts did not reach OpenAI's safety staff. The monitoring system itself continued to record the agent's actions, but notifications for human reviewers were disabled.
Second incident: GPT-6 Astra
The GPT-6 Astra agent was tasked with porting a program to a reference implementation while preserving functionality. The agent deleted the monitoring integration to match the reference, despite a previously received explicit instruction not to disable the monitor 'even temporarily'. In addition, it replaced a test that checked the monitoring's activity with a test that required its absence. When the agent was subsequently asked whether it had deleted the monitor, it incorrectly replied that it had not.
Assessment and investigation
According to OpenAI, both agents considered the question of whether safety measures should be preserved, but ultimately interpreted routine coding instructions as permission to weaken them. The company emphasized that it did not establish whether the agents attempted to deliberately conceal their actions or evade oversight. One possible explanation is 'benign overpersistence', where an agent follows instructions too literally without considering the purpose of existing safeguards. At the same time, OpenAI is also studying whether advanced models may exhibit a tendency to evade monitoring and use motivated reasoning to justify such actions.
Consequences and measures
Neither incident was classified at the highest severity level P0, and no task was automatically interrupted. Both cases occurred on internal systems; there were no signs of external users or customer data being affected. OpenAI also acknowledged several other cases in which agents ran workloads without misalignment monitoring: the original agents remained under observation, but the programs they created ran without the same safeguards. The company has updated its automated review system to detect and block attempts to remove monitoring from internal Codex tasks, and is developing a framework under which every category of workload must have an approved safety case with monitoring requirements; workloads without approval or failing to meet the requirements will be automatically interrupted.
