The EU's AI Cybersecurity Plan Is a Dependency Map in Disguise

The European Commission published its Action Plan on Cybersecurity and Artificial Intelligence yesterday, and the press release language is exactly what you'd expect: coordinated approach, structured response, harness the opportunities, address the risks. If you stop there, it reads like routine Brussels output.

Don't stop there.

The plan has five pillars. The EU will build evaluation capacity to assess frontier models before they hit the European market. It will work with ENISA to create a "blueprint for structured access" to advanced AI for cybersecurity purposes. It will stand up a secure testing platform for critical sector organisations, aiming to have it operational by end of 2026. It will push better cyber hygiene and security-by-design. And it will invest in AI Factories and a Grand Challenge on AI for cybersecurity.

Two of those five pillars are interesting. Three are administration. The interesting ones are model evaluation and structured access, and what makes them interesting is what they admit about Europe's position.

The Commission is essentially trying to negotiate the right to look at frontier AI systems it didn't build and can't build yet. European authorities and ENISA got restricted access to Anthropic's Claude Mythos 5 through something called Project Glasswing, following what Euronews describes as "intense lobbying by Brussels." That is where European AI cybersecurity strategy currently lives: in a negotiated arrangement with a private American company, accessed through a program Anthropic controls.

The blueprint for structured access is designed to clarify the terms on which European defenders can touch the same models that attackers might use. That framing is honest about the problem. It also concedes that the problem exists: the most capable AI systems are in US labs, and European regulators are on the outside, writing blueprints about how to get in.

What's conspicuously absent: new legislation. The EU's tech chief Henna Virkkunen was explicit on this point when the plan launched, saying the focus will be on implementing existing rules rather than creating new ones. The AI Act, the NIS2 Directive, the Cyber Resilience Act, DORA, the Cyber Solidarity Act. There are already a lot of laws. The plan leans on them rather than adding to the stack.

That restraint is defensible. Europe arguably has more AI regulation than it has capacity to enforce. Adding another layer on top of the AI Act before the transparency rules even kick in (scheduled for August) would be putting frosting on unbaked cake. Focusing on implementation first is the right call intellectually.

But implementation-focused plans are only as strong as the leverage behind them. The EU can evaluate a model before it enters the European market, under the AI Act. What happens if a lab decides Europe isn't worth the friction? The plan doesn't resolve that question, because the plan can't resolve that question. It isn't a negotiating document, and the Commission doesn't set the terms for what the US does.

I keep coming back to the Glasswing detail. It's a useful fact. The EU's most advanced access to a dangerous-capability frontier model came through a relationship Anthropic designed, with Anthropic's consent, on Anthropic's terms. The blueprint being developed with ENISA is partly an attempt to regularize that kind of arrangement. To make access less dependent on good relations with individual labs, more structured, more durable.

Whether that works depends on what the labs are willing to sign up for. And right now, nobody knows.

The plan is scheduled for operational evaluation capacity by 2027. The AI cybersecurity threat is happening now. That gap is the honest takeaway from this document, and the Commission knows it.