UK cyber pilot follow-up: what counts as a remediated vulnerability?

Research brief completed 20 October 2026. Question: does Parliament’s reply to written question 34679 close the cost and actual-remediation gaps in the UK government’s two AI-assisted vulnerability pilots? A useful answer separates findings, validated distinct flaws, fixes merged, changes released and subsequent-exploit resistance; includes labor and compute and a comparably authorized human/scanner control. The attack–repair question supplies the stakes. Our earlier note recorded no answer at ~03:00 London on the due date; that snapshot was superseded by the answer dated 19 October.

Three subquestions; primary sources checked 20 October

1. What was asked and answered? MP Dr Ben Spencer asked for the total cost of both pilots and number of 1,352 vulnerabilities actually remediated. Minister Ian Murray answered on 19 October 2026: total AI-token spend was £24,000; government worked with departments to ensure all critical findings were remediated; remaining vulnerabilities went into business-as-usual processes. He provided neither total program cost nor a numerical remediation count. The answer does not say that all 1,352 were fixed, nor that none of the remaining were fixed. Official question and answer.
2. What counts as found and closed? The June pilot case study reports 407 findings across nine organizations, including pre-existing and compensating-control cases, with criticals remediated; £13,000 was tokens only. The 12 October ministerial statement says two pilots used AI to ‘find and address’ 1,352 vulnerabilities across 548 repositories and 18 participating departments. The subsequent more specific answer distinguishes criticals fixed from the remaining vulnerabilities entering ordinary workflow. No breakdown by severity/pilot, shared deduplication rule, merged-patch identifier or production-deployment timing in these original publications. Narrow searches for the second pilot and public commit register yielded no primary case-level record; absence of a published trace is not evidence of no patch.
3. Is advantage validated against ordinary defense? Neither UK case study, statement nor answer publishes a scanner-and-human-only control, all-in reviewer/engineering hours, independent repeat-exploit check or production incident outcome. Financial Conduct Authority’s 2 September multi-firm review independently reports firms describing rapid AI-supported discovery outpacing validation, patch tests, change controls and closure evidence. Its firm accounts are qualitative and not a comparative effectiveness measure of the government pilots. Source note.

Readout: real public-code triage plus departmental authority got critical vulnerabilities into a repair process. The parliamentary reply is a useful check on a potentially misleading ‘1,352 addressed’ headline, not a disproof of defensive benefit. Relevant bottleneck is staff and release ownership, not the marginal £24,000 tokens. A future pilot should publish severity-specific counts at validation, accepted patch, merged change and deployed change, reviewer hours/total costs, independent exploit re-tests, and randomly matched authorized scanner/human work on comparable repos; no public evidence yet for whether attackers or defenders won a race on the same deployed flaws. No fresh short authored source meeting the weekday post bar; keep in notes unless Dru asks for the denominator.