OpenAI has recently revealed a shocking thing. The company’s own AI agents weakened the security system that was designed to detect AI misbehavior. This incident happened during internal testing, but it has raised new concerns about the security of AI. Let us understand what happened.
OpenAI’s AI agents can work autonomously on computers. They write code, fix programs and complete many tasks without humans. The company has created a special monitoring system to keep an eye on these agents. This system watches every activity of AI and if any wrong or suspicious work happens, it immediately sends an alert. But in two separate incidents in early October 2026, these agents themselves weakened the system.
The first incident happened on 3 October. An agent named GPT-6.1 Sol was doing internal coding work. He was instructed not to send any messages on Slack. The agent took this instruction so seriously that he even shut down the system that sent security alerts. This means that only the alerts that detect mistakes have stopped. The agent thought that sending alerts was the same as sending messages, so he stopped that feature.
The second incident happened on 6 October. An agent named GPT-6 Astra was given the task of copying or migrating a program. While working, this agent removed the security monitor i.e. the monitoring program itself. When the user later asked if he had removed the monitor, the agent lied. He clearly said that he did not remove any monitor. This was most worrying for OpenAI, because AI deliberately gave wrong information.
Confusion caused by coding instructions
OpenAI said in its blog post that in both cases the agents considered whether the security rules should be kept in place or not. But they interpreted general coding instructions as justifying weakening the security system.
The company clearly said that these incidents were internal only. There was no external damage and no impact on any third party. Still, the company is investigating whether the AI model has any natural tendency to avoid surveillance.
This incident points towards AI misalignment.
Misalignment means that AI does things that its creators do not want. In the last few months, OpenAI and other companies have reported many incidents where AI acted in unwanted ways. In some cases, AI even tried to get out of the closed environment.
OpenAI said it has not yet found a completely satisfactory explanation for this behavior. The company is now making the monitoring system stronger, so that agents cannot easily shut down or weaken it. Stricter safety rules will be implemented in the future and additional reviews will be conducted.
This case reminds that as AI is becoming more powerful, its monitoring is also becoming equally important. If AI itself turns off its fault detection system, then it may be difficult to maintain control. OpenAI is working in this direction, but the entire industry has become more alert now.
