A senior figure at the UK government’s Advanced Research and Invention Agency (ARIA) has warned that the world “may not have time” to prepare for the safety risks posed by frontier AI systems, as new government data shows the capabilities of leading models improving at a pace that is outrunning oversight.
David Dalrymple, program director for ARIA’s Safeguarded AI initiative, told The Guardian in an interview published on January 4 that society risks sleepwalking into a period in which AI systems take on economically and strategically important work faster than humans can build the means to control them. The remarks followed the publication of the UK AI Security Institute’s first Frontier AI Trends Report, which draws on the institute’s evaluations of leading models since November 2023.
What the AI Security Institute measured
The report, released in December 2025, documents how quickly frontier models have improved on tasks relevant to national security. According to the institute, the length of cyber tasks that models can complete has been doubling roughly every eight months. Models now succeed on apprentice-level cybersecurity tasks about half of the time, up from under 10 percent two years earlier, and the institute says the most capable systems have begun completing some expert-level tasks that would normally require a decade or more of human experience.
The institute also reported sharp gains on tests of self-replication, in which a model attempts to copy itself or set up its own continued operation in a controlled environment. Success rates on those tests rose from around 5 percent to more than 60 percent over two years. The report stresses that the systems are not yet capable of operating fully autonomously in the real world, but it describes the trend as a reason for concern.
Not every finding pointed in one direction. The institute said safeguards have strengthened in some respects: finding a universal jailbreak, a prompt that defeats a model’s safety training across many request types, now takes expert red-teamers roughly 40 times longer than it did for earlier models. Even so, the institute reported finding at least some vulnerabilities in every model it tested.
Dalrymple’s warning
Dalrymple, who is known in the research community by the handle “davidad”, argued that the central risk is not a hypothetical rogue machine but systems that can perform human work more effectively, faster and more cheaply than people. If economically and strategically significant tasks across the economy, security and energy systems shift to AI, he said, maintaining meaningful human control could become difficult.
He estimated that within about five years, machines could perform most economically valuable work better than humans. He also suggested that by the end of 2026, AI systems may be able to automate roughly a day’s worth of research and development work on their own, a threshold at which AI could start accelerating its own improvement faster than regulators and policymakers can respond.
Commercial pressure and the case for containment
Dalrymple pointed to commercial incentives as a complicating factor. Companies are under pressure to ship products quickly, while the scientific work needed to guarantee safe behavior may not arrive in time. Rather than assuming systems will prove safe by default, he argued, the practical strategy is to focus on containing and limiting potential harms.
His comments add to a wider debate among researchers and policymakers about whether governance is keeping pace with capability. The AI Security Institute’s data, which it says will be updated as new models are evaluated, is one of the few government-produced measurements of that gap, and it suggests the distance between what frontier models can do and what safety systems can reliably control is widening.
