Executive summary
AI is accelerating vulnerability discovery. Drawing on ENISA’s 2026 report, this analysis examines how leaders can strengthen patch verification, clarify approval responsibilities and prepare critical services for urgent changes.
ENISA’s 2026 report reveals a growing constraint on cyber resilience: the capacity to verify, authorise and recover from urgent changes at the speed AI makes possible.
Consider an industrial operator facing a critical software vulnerability. An AI tool has found the flaw and drafted a patch. Deploying it quickly could close the exposure before attackers exploit it; deploying it without adequate validation could interrupt the process the software controls. The operator needs to establish whether the fix is appropriate for this system, who can approve it and how service can be restored if the change fails.
ENISA’s 2026 Report on the State of Cybersecurity in the Union, published on 10 November, makes this tension explicit. Frontier AI is making vulnerability discovery increasingly industrialised, and the agency warns that triage, prioritisation, coordinated disclosure and remediation are becoming more critical bottlenecks than discovery alone. It goes further: critical service providers may be pushed towards autonomous patching because organisational approval is too slow, at the risk of updates that break systems and cause downtime, while verifying AI-generated patches may become a major bottleneck in its own right. [1, p. 48]
Read together, these passages point to a capacity constraint that deserves board attention. Faster discovery creates more opportunities to reduce exposure, but only if organisations can validate and implement the response. The delay between a credible finding and a verified, authorised intervention becomes a business risk in its own right. Reducing that delay requires engineering capacity, clear responsibilities and a way to resolve competing security and operational priorities.
Progress leaves an execution gap
The report’s overall message is one of cautious confidence. Europe’s cybersecurity posture has improved, and NIS2 acted as a catalyst for structuring and operationalising cybersecurity governance. [1, pp. 7, 62] The EU Cybersecurity Index averages 71.53 out of 100, with 22 Member States at a mature posture and two at an advanced one. ENISA revised the index for this edition, so the score is best read as a snapshot rather than a trend line. [1, p. 19]
The indicators beneath that average are more revealing. Supply-chain management practices by essential and important entities score 89.91, while participation in information-sharing communities scores 26.30 and CSIRT maturity stands at 54.76. [1, p. 20] These indicators measure different capabilities and cannot establish how effective an individual organisation’s controls are. They nevertheless show how mature practices in one area can coexist with weaknesses in the collaboration needed to manage incidents across organisational boundaries.
The same pattern appears at the level of providers. Citing its 2025 NIS Investments study, ENISA reports that 43% of surveyed entities in the ICT service management sector conducted no cybersecurity assessments in 2024, and 45% took more than three months to apply critical patches to critical systems. Because many other organisations depend on these providers, the exposure does not stay with them. [1, p. 47] A board can hold contractual assurance that a supplier has controls and still have no evidence that it can contain an exposure and restore service in time.
Faster discovery raises the demands on remediation
AI can lower the marginal cost of finding flaws and assist with assessing or fixing them. Establishing whether a proposed change is appropriate still depends on the environment in which it will run. ENISA notes that electricity, gas, railway, drinking water, waste water and public administration can depend on legacy systems that are unsupported, difficult to patch or expensive to replace. [1, p. 42] In those settings, validation must account for the safety and availability of the service as well as the vulnerability being addressed.
The report also flags an asymmetry. AI-powered discovery may increase the exposure of less-prepared critical organisations, while smaller organisations may struggle to access testing capabilities or handle the verification workload. [1, pp. 43, 48] A provider unable to validate an urgent fix can leave customers exposed even when those customers have strong internal controls.
An organisation can therefore improve detection while accumulating unresolved findings. Management should measure the full remediation cycle: validating the exposure, testing the fix, obtaining authorisation, deploying the change and confirming recovery. Measuring each interval separately helps distinguish an approval delay from a shortage of test environments, engineering resources or supplier support. The remedies differ, even when the resulting exposure looks similar.
Authority has to be designed before the incident
ENISA’s warning about slow organisational approval makes advance preparation particularly valuable. Established change-management practices offer a starting point: standard changes can be pre-approved within tested conditions; emergency procedures can bring the right engineering, security and service owners together quickly when those conditions do not apply. The practical task is to make those arrangements usable under incident pressure, with the relevant evidence ready for the people responsible.
For a critical service, that preparation includes agreed deployment conditions, monitoring, rollback or compensating controls, and a clear route for escalation when the proposed change exceeds existing approval. Where automated tools are involved, their ability to execute changes should be constrained by those arrangements. Technical evidence that a patch works and organisational authority to deploy it are separate requirements; neither substitutes for the other. ENISA’s recommendation that Member States require clear escalation paths for AI-accelerated incidents reinforces the importance of preparing these decisions. [1, p. 44]
The stakes extend beyond defensive tooling. In July 2026, Hugging Face disclosed an intrusion into part of its production infrastructure that it attributed to an autonomous AI agent system. Its disclosure describes credential theft and lateral movement, while stating that the model behind the attack was unknown. [3] In my Forbes article, “Why AI Agents Need A Chain Of Authority, Not Just A Human In The Loop,” I argued that consequential agent actions need a verifiable connection to legitimate human authority. [2] Automated remediation is a concrete setting in which that principle matters: an urgent security objective does not, by itself, establish permission to change a critical system.
Assurance has to survive the change
Automated remediation also creates a continuing assurance task. Evaluations and certificates have defined scopes, assumptions and conditions of validity. A software change may remain within those boundaries or alter something material to the assessment. Organisations need a proportionate way to determine which applies, retain the supporting evidence and obtain further assessment where necessary. This makes change-impact assessment and operational monitoring important complements to periodic evaluation.
ENISA also identifies uneven national capacity to support cybersecurity certification. The relevant indicator averages 55.56, with 13 Member States at the foundation or developing levels and 14 at the mature or advanced levels. The agency warns that uneven maturity could constrain the scalability and uptake of EU certification as more schemes become available. [1, pp. 20, 51] Capacity planning for evaluators and laboratories is therefore part of the implementation challenge.
At policy level, ENISA recommends using NIS2, the Cyber Resilience Act and the AI Act to address AI-related cybersecurity risks. [1, p. 52] For business leaders, the corresponding task is to connect responsibilities that often sit in different teams: product security, AI governance, operational continuity and supplier management. Their decisions converge whenever an automated tool proposes a change to a service on which others depend.
What boards should ask now
Three questions make that task concrete.
- How long does a verified critical finding take to become an implemented fix on our most important service, and which step dominates the delay?
- Which changes can proceed under existing approval, and what triggers escalation?
- What evidence shows that our critical providers can validate, deploy and recover from an urgent intervention? A joint exercise can expose gaps that separate policy reviews miss.
The industrial operator facing an AI-generated patch needs the scanner, the test environment, the responsible decision-maker and the recovery procedure to work as one response. As frontier AI compresses the time available to defenders, the quality of those connections will increasingly determine whether a discovered weakness becomes a repaired system or an unresolved exposure.
Sources
[1] European Union Agency for Cybersecurity (ENISA). 2026 Report on the State of Cybersecurity in the Union. Released on the EU Publications Office website on 10 November 2026. ISBN 978-92-9204-816-7. DOI: 10.2824/1662989. The report covers July 2024 to July 2026 unless otherwise specified. References use printed page numbers.
Report: https://doi.org/10.2824/1662989 | Official publication page
[2] Stéphane Gervais. “Why AI Agents Need A Chain Of Authority, Not Just A Human In The Loop.” Forbes Business Council, 11 September 2026. Read the Forbes article
[3] Hugging Face. “Security incident disclosure - July 2026.” 16 July 2026. Read the disclosure
Are you governed to decide?
Clarify decision rights, autonomy limits and investment conditions with the AI Board Decision Kit.
Open the AI Board Decision Kit →
