The Pulse
UK Expands Frontier AI Cyber Tests Across Government
The UK government will extend the use of frontier AI to find and fix vulnerabilities in public-sector systems after pilots identified more than 1,000 potential weaknesses. Ministers will also consider cyber-reporting rules for AI-related in
AI.info Team · Written with AI ·
“I want Britain to become one of the hardest countries in the world to attack using artificial intelligence.”
Kanishka Narayan, AI Minister, UK government
From pilot projects to wider public-sector use
The UK government said on 12 October it will expand work using frontier AI to strengthen cyber defences across public services. Two pilot projects have helped identify and address more than 1,000 potential vulnerabilities in government systems, while also detecting potential cyber threats, the government said.
The announcement combines that expansion with plans to assess AI-related risks to essential services, including hospitals and utilities. The government will examine whether parts of the Cyber Security and Resilience Bill regime should be strengthened, including how serious incidents and near misses involving AI are identified and reported. It has not yet set out specific legislative changes.
Scientists at the UK’s AI Security Institute are also evaluating the latest generation of advanced AI models, following work with AI developers and the US government, the announcement said. It did not name the models or publish evaluation results.
One pilot found 407 security issues
A government case study published in June offers a closer look at one strand of the testing. The Government Cyber Coordination Centre organized weekly, in-person sessions in which teams used AI to scan public code repositories across government, working with specialists from the National Cyber Security Centre and the AI Security Institute.
That project recorded 407 findings across nine government organisations over a month. The findings included weaknesses involving authentication bypass, data exposure and remote code execution; the case study says all critical weaknesses were remediated and that investigators found no evidence of exploitation. It puts the token cost of the work at £13,000.
The 407 findings belong to that June case study; the government’s October announcement gives a separate, combined figure of more than 1,000 potential vulnerabilities identified and addressed across two pilots. The June account says the next phase would add departments and models and extend the work beyond public code to closed-source systems. The October announcement does not specify when that wider rollout will begin.
Human review stays in the loop
The June pilot tested more than one method. One team used a six-stage AI-agent pipeline to examine repositories, with a person checking findings and verifying exposure; another used established scanning tools to produce leads for models to investigate. A separate team built five Claude Skills to make audits across hundreds of services repeatable.
Those safeguards reflect a key limit in the pilot’s findings: AI can surface candidate problems faster than people can validate them, and identifying a weakness does not itself fix it. Departments still had to assess findings and route exploitable issues through existing remediation processes. The case study says the strongest results came from structured workflows and expert review, not simply from choosing a particular model.
New plans address incidents involving AI agents
Alongside defensive testing, the National Cyber Security Centre and the AI Security Institute are developing plans for responding to incidents involving AI agents that can perform complex tasks with limited human supervision. The government warns that such an incident could unfold faster and affect more systems than the major incidents existing response structures were designed to handle.
Ministers will consult parliamentarians, technical experts and the operators expected to carry out any new reporting measures. The announcement also points to AI verification—checking claims about how systems were built, tested and assessed—as another area for future UK support, with more details promised in the coming months.
The immediate policy step is an examination, not a new legal duty. The government has committed to extending practical AI-assisted cyber work, but has not yet published the programme’s scale, timetable or rules for reporting AI-related incidents.