The UK government has revealed the results of an ambitious cybersecurity experiment that deployed some of the world’s most advanced artificial intelligence (AI) models to hunt for vulnerabilities across public-sector code, uncovering hundreds of security weaknesses and offering a glimpse into how AI could reshape cyber defence.
The pilot, led by the Government Cyber Coordination Centre (GC3) in collaboration with the National Cyber Security Centre (NCSC) and the UK AI Security Institute (AISI), tested whether frontier AI systems such as Claude Mythos and GPT-5.5 could identify real-world vulnerabilities in government software before malicious actors exploit them.
According to the government, the initiative forms part of the Government Cyber Action Plan, which seeks to strengthen cyber resilience across the UK public sector through emerging technologies.
“We know AI is disrupting the cyber threat landscape,” the govt said, noting that recent frontier models have delivered “a step-change in cyber capabilities” and are improving at cybersecurity tasks at a rapid pace.
Moving Beyond AI Benchmarks Into Real-World Defence
Government officials argued that traditional AI evaluations provide only a partial picture of cybersecurity performance.
“A high score on a benchmark does not necessarily translate into finding and fixing real vulnerabilities,” the government noted.
To test AI in operational settings, teams participated in weekly hackathons where they scanned public government code repositories using different AI-powered workflows. Rather than prescribing a single method, participants were given access to frontier models and allowed to build their own tools and pipelines.
Some teams created sophisticated multi-stage AI systems in which agents challenged and verified one another’s findings before a human expert reviewed the results.
Others combined conventional security scanners such as Gitleaks, Trivy, Semgrep and Hadolint with AI models capable of analyzing vulnerabilities, mapping attack chains and validating risks.
A separate government department developed five specialized “Claude Skills” designed to turn large-scale security audits into repeatable workflows that could be applied consistently across hundreds of repositories.
Hundreds of Vulnerabilities Discovered
magnific.com
The month-long pilot uncovered 407 security findings across nine government organizations, including critical weaknesses that could have exposed systems to authentication bypass, data leaks and remote code execution.
The government said some vulnerabilities were already known and mitigated through existing safeguards, while others had not been previously identified.
“All critical weaknesses have been remediated, and no evidence of exploitation was identified for any finding,” the report stated.
One of the most serious discoveries involved a legacy GitHub Actions workflow supporting a key government digital service. The flaw allowed an external user to trigger automated processes through a specially crafted comment, potentially enabling remote code execution and access to sensitive credentials.
According to the report, such access could have allowed attackers to manipulate repositories, alter contributor permissions and gain access to additional secrets within the automation environment.
AI Excels Where Traditional Tools Struggle
Officials said frontier AI models demonstrated an ability to trace vulnerabilities across service boundaries and connect business logic with technical security issues in ways conventional scanners typically cannot.
Remarkably, the entire exercise cost approximately £13,000 in AI token usage.
“The strongest results came from using frontier models as tightly scoped components inside a structured pipeline,” the report said, emphasizing that architecture and human oversight mattered more than the specific AI model being used.
Second Phase Already Planned
Despite the success, officials cautioned that finding vulnerabilities is only the first step.
“Finding isn’t the same as fixing,” the report noted, stressing that remediation, prioritization and patch management still require significant human involvement.
The government has already approved a second phase of the pilot that will involve additional departments, more AI models and expansion beyond public repositories into closed-source government systems.
Officials said the goal is to “close the gap between a theoretical benchmark and a real reduction in risk,” while ensuring AI is deployed responsibly in national cyber defence.
Leave a comment