Britain is moving frontier artificial intelligence from research labs into active cyber defense after two government pilots helped identify and address 1,352 vulnerabilities across hundreds of public-sector software repositories.
AI Minister Kanishka Narayan announced the expansion on October 12, 2026, during a speech at Imperial College London. The two pilots examined 548 code repositories across 18 government departments, while also helping organizations detect more advanced cyber threats. The government says it will now urgently extend the program and continue concentrating on weaknesses in systems considered especially high risk.
The approach combines frontier AI models with conventional cybersecurity tools and human review. In one government case study published in June, teams used AI agents to inspect public code repositories, trace attack paths and challenge one another’s findings before specialists decided whether a vulnerability was real and how serious it was. Other teams combined models with tools such as Gitleaks, Trivy and Semgrep.
That earlier month-long exercise produced 407 findings across nine government organizations and cost about £13,000 in AI token usage. Some findings were already known and controlled, while others were previously unidentified. The government said all critical weaknesses uncovered in that exercise were remediated and that investigators found no evidence any of them had already been exploited.
One example involved a legacy GitHub Actions workflow supporting a government digital service. Investigators found that a specially constructed pull-request comment could potentially trigger code execution on an automation runner and expose credentials associated with the workflow. The flaw illustrates why officials believe AI can add value beyond conventional scanners: models can follow relationships across services and business logic that simpler tools may miss.
The government is simultaneously preparing for the opposite problem — AI itself becoming a source of cyber risk. Narayan said Britain should aim to become “one of the hardest countries in the world to attack using artificial intelligence.” Officials are studying how serious incidents involving autonomous AI agents could be contained, investigated and reported, particularly when those systems operate with limited human supervision.
Those proposals are still being developed. Ministers are considering whether the Cyber Security and Resilience Bill regime should include stronger protections for emerging AI risks and mechanisms for reporting serious incidents and near misses. The bill already proposes a broader two-stage reporting system requiring covered organizations to notify regulators and the National Cyber Security Centre within 24 hours of a significant cyber incident and submit a fuller report within 72 hours.
The legislation has not yet become law. Parliament’s current record shows the Cyber Security and Resilience Bill is in the House of Lords after completing committee stage, with report stage scheduled for October 26. Many of its technical requirements would also need later secondary legislation and consultation before becoming operational.
Britain is also strengthening the institutions responsible for evaluating the most advanced AI systems. Narayan told Parliament that the AI Security Institute is testing the latest generation of frontier models, including Gemini 4 Argon, after working with developers and the U.S. government to maintain pre-release access. Separately, the National Cyber Security Centre continues to recommend secure-by-design development, detailed logging, incident response planning and continuous monitoring for AI systems.
The strategy therefore has two sides: use AI aggressively to find vulnerabilities before attackers do, while building the legal and technical capacity to respond if increasingly autonomous AI systems themselves become part of the threat. The pilots show that the defensive value is already measurable. The harder challenge will be scaling that approach across government without allowing automated findings, false positives or poorly controlled AI agents to create new security problems of their own.
