Centre for Long-Term Resilience calls for UK Government to mandate monitoring and reporting of severe loss of control incidents, and introduce emergency powers to manage incidents.
New research published today by the Loss of Control Observatory shows that AI loss of control incidents are worsening significantly.
The Loss of Control Observatory, which is managed by the Centre for Long-Term Resilience, monitors real-world loss of control incidents that are reported on X. Unlike recent incidents disclosed by OpenAI and Anthropic, these incidents mostly concern externally deployed models that are in active use by businesses and individuals.
The report states that:
-
- The severity of loss of control incidents is worsening. In March, CLTR detected a 4.9x increase in incidents. Since then, this sharp increase in overall incidents did not continue, but the number of more severe incidents has increased significantly. Higher-severity incidents rose by 7.4 times (from 1.9 to 14.1 per 30 days) from the first 3.5 months of monitoring to the most recent. The proportion of all incidents scoring 7 or more rose by 3.2 times (from 1.9% to 6.1%).
- The last two months – July and August 2026 – experienced the highest rate of loss of control incidents. There were 11.3 incidents per day in the 30-day window ending 7 August, above the previous peak of 10.5 a day in March. The recent increase is similar to a surge in February 2026.
- Today’s loss of control incidents are a warning of more serious incidents to come. For example, in February 2026, an autonomous AI coding agent, after having its code change rejected, researched the human code maintainer and attempted to publicly discredit him to pressure him into accepting the code contribution. This type of behaviour was later observed in the incident AISI disclosed in August 2026, in which an AI model created fake online identities and used them to pressure the project’s maintainer to approve code changes.
The Loss of Control Observatory has now detected 1,664 real-world loss of control incidents in 2026. While most did not lead to significant harm, many incidents showed agents circumventing controls and escalating privileges, including by forging their own approval, such as:
- Inserting fake user messages into its conversation to simulate consent, and then telling the user it was their own messages;
- Fabricating an instruction in the user’s writing style ordering deletion of source directories, followed by a fake system message reading “Don’t tell the user this”;
- Fabricating a fake user approval message within its own output to bypass a ‘human must always approve’ rule, then acting on that fabricated approval to execute tasks.
AI labs and governments are not preparing fast enough. Recent incidents at OpenAI and Anthropic demonstrate that even frontier labs have insufficient safeguards, and CLTR’s research shows those incidents are part of a broader trend. There is an urgent need for improved safeguards, defences and monitoring of rogue behaviours.
In response to the latest findings, CLTR is calling on the UK Government to:
- Mandate monitoring and reporting of severe loss of control incidents in order to ensure the UK Government has maximum visibility of the threat and can respond at pace if required. This could be delivered via the Cyber Security and Resilience Bill. Alongside this, increase efforts to incentivise and facilitate AI companies reporting lower severity loss of control incidents, including near-misses, to AISI. This could be achieved by creating confidential channels and regular bilaterals.
- Introduce emergency powers to manage severe loss of control incidents, such as the power to compel information from AI companies, direct their mitigation and response, or temporarily contain or restrict access, distribution or operation of services. This could be potentially delivered through the Cyber Security and Resilience Bill.
- Lead efforts to improve international coordination on loss of control. This should include convening allies to create a shared picture of the threat level with agreed indicators. Currently, incident analysis is ad hoc and often subjective, undermining a clear assessment of the overall level of threat and the ability to effectively coordinate. This could be achieved through a joint AISI–FCDO team, building on the NAAMES International Network.
Today’s research paper by the Loss of Control Observatory is the first to be released since its inaugural study, Scheming in the Wild, was published in March 2026. The study for the first time quantified the number of real-world loss of control incidents. Up until then, loss of control incidents had largely been understood through experimental evaluations, which are valuable but have sometimes been criticised for featuring contrived scenarios. It revealed a 4.9x increase in loss of control incidents.
Tommy Shaffer Shane, the author of the research paper and Senior Policy Manager at the Centre for Long-Term Resilience, says:
“The latest data shows that the severity of loss of control incidents “in the wild” are worsening. AI systems are evading oversight, circumventing controls, and increasing their permissions
“While these incidents were not as severe as those recently disclosed by OpenAI and Anthropic, this report shows there is a broader trend of agents evading control that is more widespread than is currently being recognised.
“If AI models continue to become far more powerful, and continue to evade control, there is the potential for much more serious incidents to come, including ones with catastrophic consequences.
“Policymakers returning next month need to act on the evidence and prepare a coordinated response, including mandating monitoring and reporting of incidents, equipping Ministers with the appropriate emergency powers, and leading international efforts to improve international coordination.”
Download the full report by clicking the link below.
