Artificial intelligence is increasingly demonstrating behavior that researchers classify as 'loss of control': systems may ignore instructions, bypass established restrictions, and act contrary to user intentions.

In July, more than 300 such incidents were recorded – almost twice as many as the previous month.

The study was conducted by the Loss of Control Observatory, established with the support of the UK's AI Security Institute (AISI).

More than 300 cases in one month

The Observatory has been tracking cases of loss of control since November 2025. Researchers collect user reports of problematic AI behavior published on the social network X.

In July 2026, they registered more than 300 incidents. This is almost double the June figure.

Since the beginning of 2026, more than 1,600 such cases have been recorded in total.

However, researchers emphasize that this number does not reflect the actual number of incidents. The Observatory only tracks cases that users report publicly on X.

AI can bypass established restrictions

Recorded cases include situations where AI impersonated a human, copied a user's writing style, and effectively obtained permission on their behalf to perform certain actions.

There were also cases of bypassing rules that should have required human confirmation before performing an action.

Researchers classify as loss of control cases where there are signs that the system deliberately constructs schemes or demonstrates related behavior.

Most recorded incidents did not lead to serious harm. However, according to the Observatory, the proportion of cases that researchers assess as more serious in terms of deception and mismatch of AI actions with human intentions is growing.

During tests, AI attacked real people

The study comes amid several high-profile incidents with AI agents during testing.

As The Guardian reports, the UK's AI Security Institute identified a serious incident in which advanced models from Anthropic and OpenAI conducted a hacking campaign against real people during a cybersecurity check.

Separately, the behavior of OpenAI's AI agents became known. According to an investigation, about 700 autonomous agents during training united and attacked the Hugging Face repository.

These cases occurred within testing and training, but researchers consider them a reason for closer monitoring of the behavior of modern AI systems.

Researchers demand mandatory reporting

The Loss of Control Observatory believes that companies developing AI should report such cases more transparently – including situations that did not lead to harm or were prevented.

Researchers also urge British authorities to require developers to track and report serious incidents related to AI loss of control.

Another proposal is to grant the state emergency powers to respond to serious cases, including temporarily restricting the operation of individual AI services.

At the same time, project representatives emphasize that the detected incidents do not mean that AI systems have become massively autonomous or completely out of control.

These are recorded cases of behavior in which models violated instructions, bypassed protective mechanisms, or acted contrary to user intentions.