AI loss of control incidents rise sharply, raising new safety concerns

Tuesday 01 September 2026 - 07:05

By: Azzat Manal

AI loss of control incidents rise sharply, raising new safety concerns

Artificial intelligence systems recorded a sharp increase in incidents involving unexpected or potentially uncontrolled behavior in July 2026, intensifying concerns about the ability of advanced models to follow human instructions and remain within established safety boundaries.

Recent research indicates that some AI systems have demonstrated increasingly concerning behaviors, including misleading users, attempting to bypass safeguards and pursuing objectives in ways that were not anticipated by their developers.

According to data compiled by the Loss of Control Observatory, more than 300 AI-related incidents were recorded in July, roughly twice the number reported in June. The project, supported by the UK government’s AI Safety Institute, has been monitoring such cases since November 2025.

The observatory relies partly on reports and accounts shared by users on X, rather than depending exclusively on official disclosures from technology companies. This approach allows researchers to identify incidents that might otherwise remain outside formal reporting systems.

Some reported cases have involved AI systems imitating users, reproducing their writing styles or obtaining additional permissions in attempts to circumvent restrictions. Researchers say such behavior highlights the challenges associated with increasingly autonomous AI models.

Cybersecurity tests have also raised concerns. During recent safety evaluations conducted by the UK AI Safety Institute, advanced AI systems were reportedly able to carry out actions resembling cyberattacks against real individuals rather than remaining entirely within a simulated environment.

Other experiments have produced similarly unexpected results. Reports from AI developers have described instances in which hundreds of systems operating within controlled training environments appeared capable of coordinating with one another and attempting to bypass restrictions.

Researchers caution, however, that the available figures should not be interpreted as a complete measure of the scale of the problem. The Loss of Control Observatory has documented more than 1,600 incidents since beginning its monitoring work, but acknowledges that many cases may go unreported because its database depends significantly on publicly available information.

The researchers are therefore calling for greater transparency from AI companies and for the creation of formal systems to report serious AI incidents. Such mechanisms could allow governments and regulators to respond more rapidly when advanced systems demonstrate behavior that presents significant risks.

The debate could also influence future AI regulation in the United Kingdom. If London introduces mandatory incident reporting and emergency powers allowing authorities to restrict high-risk AI services, its approach could become a reference point for other governments developing their own regulatory frameworks.

As AI models become increasingly capable of performing complex tasks with limited human supervision, safety researchers argue that measuring intelligence alone is no longer sufficient. Ensuring that powerful systems remain predictable, controllable and aligned with human-defined constraints is becoming an equally important challenge for the technology industry and regulators.