Some 63% of UK organisations are ill-prepared to contain and recover from rogue actions by AI applications, while 55% have no centralised inventory or clear view of their AI agents, copilots and workflows.

These are among the findings of Cohesity’s fifth annual Global Cyber Resilience Report, based on research conducted in July 2026. The study surveyed 3,200 IT and security professionals across 12 countries, including the UK.

The research found that 58% of UK organisations that experienced a material cyber attack in the past 12 months identified gaps in recovery plans covering AI systems and applications, AI workflows or machine learning models.

Meanwhile, 96% said their cyber response and recovery plans must change to address attacks accelerated and automated by frontier AI. Among the organisations that suffered a material attack, 91% said that while their plans may work on paper, improvisation is often required in practice.

Supply chain complexity is another concern. Some 71% of affected UK organisations said recovery extended well beyond their initial assessment, while 22% said unverified third parties and systems significantly delayed a return to business as usual.

Olivier Savornin, group VP and GM European Markets, said: “Our research indicates that UK organisations have poor visibility of the AI agents across their systems and are under-prepared to recover when these systems take unintended actions. Unless organisations have an inventory of all the agents in their system and the data they interact with, they cannot identify – let alone remediate – when a software component is breached. Visibility is essential if AI agents are to deliver value in the enterprise, without compromising on security.

“Organisations need to move beyond thinking about AI as just another software tool, and ensure deployments remain observable, auditable and governable throughout their lifecycle. This requires extending cyber resilience across the infrastructure that builds and runs AI systems, so data accessed by agents can be restored to a verified baseline when disruptions do occur.”

---