UK to Expand AI Vulnerability Hunting: 407 Flaws for £13,000

The UK government will expand a program in which frontier AI models search for vulnerabilities in the state's own systems. In one month, the pilot uncovered 407 weak points across nine government organizations. Tokens cost about £13,000. Alongside the expansion, the government will consider legislative measures: mandatory reporting of AI incidents and regulation of the risk of losing control over AI systems.

What the pilot showed

The pilot's numbers are easy to work out. £13,000 across 407 findings comes to about £32 per vulnerability. Each of the nine organizations accounts for an average of about 45 findings and roughly £1,440 in costs. For a government IT landscape, where security reviews are usually measured in person-days, that is a very low entry price.

Caveats apply. The available portion of the material does not disclose which models were used or how the findings were distributed by severity. It is also unknown how many findings engineers confirmed. The average cost cannot be compared with the price of a critical vulnerability. Some of the 407 records are almost certainly minor configuration issues. Still, the sheer scale shows that models are suited to broad first-pass scanning.

Two parts of the package

The package consists of two measures that differ in nature.

The first is operational: expanding the program. The state is turning an experiment into a working tool and pointing frontier models at its own systems. It is the most direct way to test what the technology can do before someone else does.

The second is regulatory, and its wording is softer. The government will "consider" legislation on AI incident reporting and on the risk of loss of control. There are no obligations yet, only an intention to study the issue. But it was placed in the same package as cybersecurity. The logic is readable: if a model can find holes in government systems, then failures of the models themselves should also come to the regulator's attention. The experience of other sectors, such as mandatory notification of cyber incidents, shows that statistics cannot be collected without a single reporting format.

The economics of defense

The pilot's main takeaway is economic. If finding vulnerabilities costs tens of pounds per finding, the same math holds for attackers. Access to frontier models is not a state privilege, which means the window between a flaw appearing and being exploited is narrowing for everyone. Defenders are left with one option: find flaws earlier and more cheaply than the adversary does.

Key parameters of the pilot:

• 407 vulnerabilities found in one month;
• 9 government organizations in the sample;
• about £13,000 in token costs;
• about £32 per finding;
• an average of 45 findings per organization.

What security teams should do

Government and corporate teams should repeat the calculation on their own data. To do so, take a limited perimeter and a month of time, then measure three things: the share of confirmed findings, the cost of verifying each one, and the false-positive rate. Verification, not tokens, will most likely be the main expense.

Second, work out in advance how to record incidents involving AI systems. If reporting becomes mandatory in the UK, companies that already have a format for logging failures will lose less time adapting. Third, follow the wording of the future bill on loss of control. It is still unclear which systems it will cover and what thresholds will define "frontier" models. The answer determines whether the regulation will affect only the labs or also those who embed their models in products.