Summary
- ENISA says frontier AI is compressing the vulnerability management lifecycle from discovery to exploitation.
- The agency warns of heavier pressure on patching, legacy systems, open-source maintainers, SMEs, and detection teams.
- Its recommendations point towards AI-resilient defences built on security fundamentals, segmentation, telemetry, and faster coordination.
ENISA has warned that frontier AI models are starting to put established cyber defence assumptions under strain, particularly by accelerating the process from vulnerability discovery to exploitation.
In a new paper, ENISA’s view on cybersecurity in the frontier AI era, the European Union Agency for Cybersecurity says national authorities, EU policymakers, defenders, manufacturers, and service providers need to prepare for machine-speed threats while keeping basic controls at the centre of resilience work.
The agency’s analysis focuses on the vulnerability management lifecycle and the attack chain, from reconnaissance and exploit development to lateral movement and response. ENISA says the frontier model landscape is evolving rapidly and expects open-weight models to approach similar capability levels within 9 to 12 months, while existing models combined with skilled security experts can already produce strong results.
The risk is not limited to highly specialised attacks. ENISA says AI may allow attackers to access exploits before fixes are released, creating a negative time-to-exploit problem. It also warns that AI could intensify challenges around legacy systems and end-of-life products, increase patch release frequency, overload open-source maintainers with vulnerability reports, and create particular pressure for small and medium-sized businesses that may lack access to advanced models and guidance.
Pace is the central operational issue. Security functions already struggle with asset inventories, patch prioritisation, exposed internet services, supplier dependencies, and fragmented telemetry. When vulnerability discovery and exploit generation accelerate, those existing weaknesses become less forgiving. An organisation that takes weeks to understand exposure to a widely used component faces a different risk calculation when attackers can validate and operationalise new flaws in hours or days.
ENISA avoids treating AI as a replacement for security fundamentals. Its paper reinforces asset management, vulnerability management, access controls, segmentation, monitoring, incident response, and supply chain security. The agency also points to zero trust segmentation, behavioural baselining, and detection layers that assume attackers will adapt in near real time.
Governance will determine whether those controls work under pressure. AI-enabled vulnerability pressure will force decisions about downtime, emergency change, supplier readiness, patch testing, and risk acceptance. Those decisions cannot sit only in ticket queues, particularly in regulated sectors operating under NIS2, the Cyber Resilience Act, DORA, and national critical infrastructure regimes.
Open-source security sits close to the centre of the problem. If AI produces more vulnerability findings, maintainers may face higher reporting volumes, heavier triage workloads, and more pressure to distinguish meaningful reports from noise. Enterprises and public bodies will still depend on the same libraries, packages, and components, but many will have limited visibility into maintainer capacity or remediation timelines.
ENISA says its recommendations are not a complete checklist and will be refined with member states and EU bodies. Alignment with the European Commission’s AI and cybersecurity action plan will be important because Europe is trying to manage AI cyber risk across regulation, operational coordination, certification, critical sectors, and industrial policy at the same time.
The paper gives European authorities a common vocabulary for what is changing. AI is altering the economics and tempo of vulnerability discovery, patching, detection, and response. Organisations that treat the issue only as AI governance may miss the infrastructure and operating discipline that will determine whether critical services can withstand a faster threat cycle.
