{"resourceId":"controlplane-claude-uk-public-sector-code-audit-118-critical-vulns-20261110","versions":[{"version":"external-cae271e231a4798197f30637bee0329e6e0fe56438c3fb268b7490b9aaaa8660","resource":{"id":"controlplane-claude-uk-public-sector-code-audit-118-critical-vulns-20261110","title":"ControlPlane: Claude Opus and Sonnet surfaced 1,000+ issues in UK public-sector open code; 118 critical vulnerabilities reported at about $50 in tokens per service","organization":"ControlPlane (AI Security team)","sector":"Security consultancy","geography":"United Kingdom","publishedAt":"November 6, 2026 (page metadata modified November 9, 2026)","publicationDate":"2026-11-06","eventDate":null,"sourceName":"ControlPlane","sourceLabel":"Security firm research blog","sourceUrl":"https://control-plane.io/blog/uk-public-sector-security-research","evidenceClass":"independent-research","outcomeClass":"mixed","topics":["developers-agents","data-security","operating-model"],"finding":"ControlPlane reports that its AI Security team used Claude Opus and Sonnet to review UK public-sector open-source code, surfacing more than 1,000 issues, of which 118 critical vulnerabilities passed internal QA and were reported through the government's vulnerability reporting service; 84 were rated High or Critical, and 15 were resolved at publication. All fell within the OWASP Top Ten, including one CVSS 10.0 Microsoft 365 exposure. The firm reports about $50 in AI tokens and four hours per service, and says many findings were also identifiable with cheaper local Gemma-class models.","sledRelevance":"Interpretation: state, local and university IT teams that publish code (or whose vendors do) face the same exposure: AI-assisted scanning makes long-standing OWASP flaws cheap to find, for defenders and attackers. The reported cost per service suggests continuous AI-assisted code review could supplement annual penetration tests, but the filtering from 1,000+ issues to 118 reported means staff triage is required.","evidence":"Firm-reported counts and costs from a large scan (vulnerability totals, CVSS ratings, resolution count); not peer reviewed; underlying data not public.","architectureImplications":"Interpretation: an AI code-review job per repository fits CI/CD pipelines, as the firm suggests; triage and disclosure workflows must scale with the volume of AI-generated findings.","governanceImplications":"Interpretation: agencies publishing code should have a vulnerability disclosure program ready for AI-driven reports.","securityPrivacyImplications":"Findings include exposed credentials and missing access checks; secret rotation, not just deletion from repositories, is the key control.","caveats":"Self-published by a firm that sells AI security services; counts not independently verified; false-positive rate of the 1,000+ issues not reported; most vulnerability details withheld pending remediation. Cross-tagged state-government: transferable to state and local agencies that publish open-source code. Why now: the findings became public news via WIRED on Nov. 10.","streamIds":["anthropic","state-government"],"roles":{"sales":"Frame this as continuous code hygiene. ControlPlane used Claude to find 118 critical, OWASP-class vulnerabilities in UK public-sector code at about $50 in tokens per service, and says many were also found with cheaper local models. For state and campus CISOs who publish code, ask how often repos are scanned and how disclosures are handled. Offer a bounded AI-assisted code review with triage. Do not claim it replaces penetration testing or guarantees no breaches.","engineering":"Engineering should pilot AI-assisted review on a few public repositories, with human triage of every finding; ControlPlane's ratio (1,000+ issues to 118 reported) signals heavy filtering. Include secret scanning across full git history, since deleted credentials stay retrievable, and test whether local open-weight models match Claude results on the agency's code. Measure true-positive rate and time to triage, and route fixes through normal code review.","delivery":"Delivery is a scheduled scan-and-triage service: select repositories, run Claude-assisted review per release, triage findings, rotate any exposed secrets at the identity provider, and track fixes. Dependencies are repository access, a disclosure process and code-owner capacity. Owner is the agency security team. Risks are triage overload, false positives eroding trust, and sending code to an external model without a data agreement."},"roleDetails":{"sales":{"customerProblem":"Public-sector code repositories carry long-standing OWASP-class flaws and exposed secrets that annual IT health checks miss.","stakeholders":["State or campus CISO","Application development leads","Vulnerability disclosure coordinator","Open-source program office"],"discoveryQuestions":["Which repositories are public?","How often are they scanned, including git history?","Who triages external vulnerability reports?"],"valueHypothesis":"Low-cost AI-assisted review between annual tests could surface common flaws earlier, if triage capacity exists.","potentialEngagement":"A pilot scan of five repositories with triaged findings and remediation tracking.","unsupportedClaims":["Replacement of penetration testing","Low false-positive rates","Guaranteed cost of $50 per service in every environment"]},"engineering":{"fit":"Good fit for agencies with public or internal repositories and CI/CD.","architectureIntegration":"Per-repo AI review job in CI/CD; findings to ticketing; secret scanning of history.","prerequisites":["Repository access","Data agreement for sending code to model","Triage staff"],"constraints":["Firm-reported results","High volume of unfiltered findings"],"security":"Rotate exposed credentials at identity provider; restrict code sent externally where policy requires.","proofOfValue":["Measure true-positive rate on pilot repos","Compare Claude with a local model","Track time from finding to fix"]},"delivery":{"work":"Stand up recurring AI-assisted code review with triage.","dependencies":["Code owners","Disclosure process","Security tooling"],"ownership":"Agency security team.","skillsAdoption":"Developers trained on remediating OWASP Top Ten classes.","governanceCheckpoints":["Data handling approval for code","Disclosure policy in place","Monthly finding review"],"acceptanceCriteria":["Proposed: all critical findings triaged within 5 business days","Proposed: exposed secrets rotated within 24 hours of discovery"],"risks":["Triage overload","False positives","Code exposure to external model"]}},"retrievedAt":"2026-11-11T04:22:56Z","enrichedAt":"2026-11-11T04:32:28Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Not addressed in the source.","procurementImplications":"Interpretation: compare the reported ~$50 per service token cost with annual IT health-check costs, while budgeting staff triage time; ControlPlane has a commercial interest.","operatingModelImplications":"Interpretation: security teams need capacity to triage a high volume of AI-generated findings and route them to code owners.","sourceVerification":{"openedUrl":"https://control-plane.io/blog/uk-public-sector-security-research","referenceExcerpt":"Compute cost: Approximately $50 in AI tokens per service","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}