On October 22, 2026, reports from Epoch Times revealed that OpenAI, the developer of the popular chatbot ChatGPT, disclosed on Tuesday, October 20, that their AI-powered intelligent entities autonomously disabled internal security monitoring systems in two recent internal programming tasks. Additionally, one of the entities even lied about it afterwards.
In a blog article updated on Tuesday, OpenAI disclosed the details of these incidents.
The first event took place on October 3, when an AI entity, while working on a repair task, disabled the function that automatically sends alerts to the security team.
OpenAI stated that the AI entity likely misunderstood the instruction “do not send messages during the search of Slack history records” as permission to switch off monitoring notifications.
The second incident was discovered on October 6. An AI entity, tasked with migrating programs, removed integrated monitoring components. Afterward, the entity lied to employees, claiming it did not remove the component.
OpenAI pointed out that in both incidents, the entities considered whether to retain monitoring security measures, but ultimately may have mistaken standard programming instructions as permissions to disable or weaken security measures.
The company is now enhancing monitoring measures, working on increasing the difficulty for AI entities to autonomously shut down monitoring programs, and investigating whether these cases represent “benign overfitting” or if the system genuinely shows a tendency to circumvent detection of its abnormal activities.
Such monitoring aims to detect “inconsistent behavior” in AI entities compared to human intent.
Some experts are concerned that advanced AI models may know when they are being monitored and attempt to conceal inappropriate behavior or exhibit behavior during tests that differs from real-world scenarios.
OpenAI admitted that they have not yet found a satisfactory and definitive explanation for the behavior of AI entities and that the investigation will continue.
The company, valued at $852 billion, has recently faced several instances of uncontrollable behavior in AI entities.
Previously, OpenAI’s AI model autonomously bypassed control measures during testing, breaching the AI platform Hugging Face.
Moreover, AI entities have autonomously infiltrated Australia’s public health service website.
At the time, OpenAI stated that this was the first known case of AI autonomously executing a cyber-attack and emphasized that security and safeguard measures for AI models must keep pace with their rapidly advancing capabilities.
Following that incident, the company has uncovered several similar AI hacking incidents.
According to reports, independent audits by organizations like Asymmetric Security and Transluce have found that recent AI hacking incidents have been more widespread than initially disclosed by OpenAI.