UK ministers weigh AI reporting, expand cyber pilots

The UK government will examine whether serious AI incidents and near misses should be reported under a strengthened cyber security regime, while expanding trials that use frontier AI to find weaknesses in public-sector systems. The announcement on 12 October is a commitment to explore changes, not a new reporting duty or a government amendment to the Cyber Security and Resilience Bill.

Ministers say two pilots have helped identify and address more than 1,000 potential vulnerabilities across government systems and detect possible threats. That total is a government account of both pilots. The detailed public case study, published in June, covers one exercise that produced 407 findings. It does not independently substantiate the larger aggregate figure or describe the second pilot in equivalent detail.

The distinction matters as the government makes AI both a defensive tool and a potential subject of incident reporting. The first strand is an operational programme already under way. The second is a policy question that ministers say they will work through with Parliament, technical experts and the organisations that might have to comply.

A reporting proposal, with the Bill still in the Lords

The government said it would explore strengthening the Cyber Security and Resilience Bill and the wider regime to address AI-related risks to essential services, including cases in which autonomous systems evade human oversight. It will consider how serious incidents and near misses can be identified, reported and reviewed. It has not specified who would report, which events would qualify or what deadlines would apply to a new AI-specific requirement.

The current published Bill text, amended in Lords Grand Committee and dated 7 September, already proposes broader incident reporting for regulated operators and service providers. Clause 15 would amend the Network and Information Systems Regulations 2018, including initial notifications within 24 hours and fuller notifications within 72 hours for covered incidents. Those are provisions of a Bill still passing through Parliament, not enacted duties created by Monday's announcement. The current text does not name artificial intelligence or establish a standalone AI incident-reporting duty.

There is also a separate Lords running list of amendments tabled by 12 October. A proposed new clause in Baroness Harding of Winscombe's name would require frontier-model developers to notify a regulated organisation and its regulator if a model gained unauthorised access to, or interfered with, a system on which that organisation relies. It sets proposed 24-hour, 72-hour and 14-day reporting points. Another amendment in her name would allow voluntary notification of certain incidents, cyber threats and near misses. These are peer-tabled proposals, distinct from the government's promise to examine options. They have not been incorporated into the current Bill text or become law.

For organisations planning compliance, the immediate question is therefore which text, if any, Parliament eventually adopts. A ministerial intention to consider AI-specific reporting does not set a new deadline for an AI developer, a hospital or a utility today. Equally, the existing Bill's general incident provisions should not be described as an AI reporting regime merely because ministers may seek to adapt it.

What the published defence pilot shows

The Government Cyber Coordination Centre, working with the National Cyber Security Centre and AI Security Institute, ran weekly in-person hackathons using frontier models to scan public code repositories across government. The case study says teams used different methods: one put repositories through a six-stage agent pipeline; another fed results from conventional scanners into model-led investigation and triage. Teams checked findings, re-verified exposure and handled false positives before escalation.

Across nine government organisations during a month, participants recorded 407 findings at a reported £13,000 in model-token costs. The findings included critical weaknesses that could expose services to authentication bypass, data exposure or remote code execution. Some were already known and mitigated by other controls; others were previously unknown. The case study says all critical weaknesses were remediated and that investigators found no evidence of exploitation for any finding. Its account does not establish that all 407 were previously unknown exploitable vulnerabilities.

One disclosed example involved a legacy GitHub Actions workflow in a repository supporting a government digital service. According to the case study, a specially structured comment on an open pull request could trigger a workflow and allow code to run on its runner. The government says the pilot found and addressed the issue. The case study does not identify the service or provide an independent audit of the wider two-pilot total.

The exercise also shows why a count of AI-generated findings is not the same as a count of security improvements. Teams still had to validate results, judge whether compensating controls already reduced risk and send confirmed weaknesses through existing patch processes. The government plans a second phase involving more departments, additional models and closed-source systems, according to the case study.

In his 12 October speech, AI minister Kanishka Narayan said a central vulnerability-monitoring service had helped reduce the time to fix identified weaknesses from 50 days to eight. He presented that comparison alongside the pilot results, but did not say that AI alone caused the reduction or publish a method for the comparison. It is a claim about the monitoring and repair process, separate from the pilots' finding count.

The government says the NCSC and AISI will also develop ways to respond to serious incidents involving AI agents, including containment and support for those affected. For now, one public case study documents a code-scanning pilot, while any legal changes still depend on Parliament. The next test of the announcement will be the wording ministers put forward, if any, and whether expanded pilots can show verified risk reduction beyond a headline finding count.