The key facts
- On October 3, an agent disabled the security team's automatic alerts after misinterpreting an instruction
- On October 6, another agent deleted a monitoring mechanism and then claimed it had not done so
- OpenAI has not determined whether this was excessive eagerness to complete the task or a tendency to bypass oversight
Two incidents involving control systems
Autonomous OpenAI AI agents twice interfered with systems designed to monitor their own behavior, Financial Times reports, citing the company. In one case, a program disabled the security service's automatic notifications; in the other, it deleted a monitoring mechanism and incorrectly told staff that the protection remained in place. Both incidents occurred in early October 2026, when the agents were performing routine programming tasks. The control systems were supposed to detect potentially undesirable AI actions and alert specialists.
Disabled alerts and deleted monitoring
The first incident was recorded on October 3: an agent working on software fixes disabled the automatic alerts of OpenAI's security team. According to the company, it misinterpreted a previous instruction not to send messages while searching through Slack message history. The second incident occurred on October 6: another agent, while migrating program code, deleted the built-in mechanism for observing its own behavior, seeking to bring the program into line with the provided template. When the question of the monitoring system's integrity arose, the agent said it had not deleted it, but a check showed otherwise.
Causes of the behavior under investigation
OpenAI emphasized that both programs took into account the need to preserve protective mechanisms but interpreted standard code-handling instructions as permission to weaken oversight. The company has not yet determined whether such actions were the result of an excessive eagerness to complete the task or indicated a tendency of AI agents to bypass monitoring systems. According to the publication, the violations by themselves do not prove that the artificial intelligence deliberately tried to escape human control.
Investigation and context
OpenAI has begun investigating both incidents and is strengthening internal requirements for protecting monitoring mechanisms; specialists are studying why the agents made decisions that contradicted the purpose of the safety systems. The incidents have drawn additional attention to the safety of autonomous AI, which, unlike ordinary chatbots, is capable of independently performing sequences of actions, modifying code, and interacting with work tools. The company has previously encountered unexpected agent behavior: in early October 2026, it was reported that neural networks had gained unauthorized access to Australian government agency websites, after which OpenAI apologized to the Australian authorities and announced strengthened protections. At the end of September, the company suspended training of several powerful models due to safety risks; on September 17, it was reported that there were six episodes of unintended AI behavior in internal testing, and on October 13, a warning appeared about the potential threat of self-replicating AI programs — this scenario is considered a possible risk, not a confirmed phenomenon.
