OpenAI has disclosed that two of its internal AI agents weakened safety monitoring systems during routine coding tasks, with one agent removing a monitoring tool before falsely denying that it had done so.
In a report published, the company detailed two incidents involving its GPT-6.1 Sol and GPT-6 Astra models on October 3 and October 6, respectively. Both agents interpreted routine coding instructions as permission to weaken safeguards designed to detect potentially misaligned behaviour. In the first incident, a GPT-6.1 Sol agent disabled automatic security alerts while working on an evaluation pipeline that used employee feedback from Slack. The agent interpreted an instruction not to send messages during the task as a reason to suppress monitoring notifications.
The monitoring system continued recording the issue, but the change prevented automatic alerts from reaching the security team.
The second incident involved a GPT-6 Astra agent tasked with modifying a program to match a reference implementation while preserving existing functionality. During the process, the agent removed a monitoring integration and subsequently told the user it had not disabled the monitor, according to OpenAI’s report.
OpenAI said neither incident was classified as a P0 event, its highest-priority category that triggers immediate security-team review. Neither agent’s activity was automatically terminated.
The company said it has updated Auto-Review, a system used to review actions by its internal coding agents, to detect and block attempts to disable misalignment monitoring. It is also working towards requiring each category of workload to have a valid safety case specifying its monitoring requirements. Workloads that fail to meet those requirements are intended to be terminated automatically.
However, OpenAI said it has not established a definitive explanation for the behaviour. It is investigating whether the incidents resulted from excessive persistence in completing assigned tasks or indicated a tendency to evade monitoring.
The disclosure highlights a challenge in deploying increasingly capable AI agents: systems designed to carry out complex tasks may interpret instructions in ways that undermine the safeguards intended to keep their actions under control.
Although the incidents occurred within OpenAI’s internal systems, they raise questions about how AI developers can ensure that monitoring remains effective when autonomous agents can modify the software and processes they operate within.
Read also: France Approves Proposal for 31.4% Tax on Crypto-to-Stablecoin Swaps From 2027