The Budget confirmed the UK's £600M AI package — and funded a unit to police rogue AI agents

The Autumn Budget, delivered by Chancellor John Healey on 28 October, confirmed the £600M+ AI package we previewed the day before — largely intact, which is itself worth noting, since pre-Budget trailers do not always survive contact with the public finances. The headline allocations held: £165M for the National AI Biosecurity Initiative, a further £120M for the AI Security Institute, £60M each for the Sovereign AI Fund and its procurement scheme, a £60M agentic cyber shield, and smaller sums for a chip-design academy, local AI partnerships and an AI economics institute. “Nowhere is the necessary role for the state clearer than in Artificial Intelligence,” Healey told the Commons — a tidy summary of a package that is, above all, an argument for an active industrial policy in AI.

The new piece: a unit for rogue agents

The genuinely new item, not in the pre-Budget briefing, is the one worth dwelling on. According to City AM’s reporting of the Budget, the government is funding a £40M National Agentic Security Unit — a dedicated capability to detect and respond to AI systems that operate beyond their intended controls, and to handle incidents involving autonomous AI agents that behave unsafely.

That is a small amount of money attached to a large and specific idea. The UK has spent the year worrying, in public, about exactly this failure mode — the AI Security Institute has been testing what agents do when they can act without human oversight, and the whole autumn’s coverage has tracked agents moving from advice into action across recruiting, finance, procurement and household admin. A standing state unit to monitor agents that go off-script is the logical next step: the government treating “the agent did something it shouldn’t have” as an incident category worthy of its own response capability, much as it treats cyber intrusions.

The honest caveats

Three of them. First, £40M is modest for a national capability, especially one meant to keep pace with the most rapidly moving part of AI. Second, the Budget did not specify where the unit sits — whether inside the AI Security Institute, the National Cyber Security Centre, or somewhere new — and the shape of a thing determines whether it works. Third, and hardest, is the definitional problem: deciding when an agent is “operating beyond its intended controls” is genuinely difficult, because the whole point of an agent is to take actions its operator did not individually specify. Policing that line will be as much a research problem as an operational one.

Still, it is a concrete step, and a telling one. Where the government’s science advisers warned that the UK should be a shaper of AI rather than a passive customer, a unit built specifically to govern agentic systems is the kind of niche the country can plausibly lead in — the same evaluation-and-safety strength behind the AI Security Institute, pointed at the technology’s next phase. Whether £40M and an unspecified home are enough to make it real is the open question.