The EU received the software bug-hunting model later. It may only use it for defense

The European Union Agency for Cybersecurity, ENISA, has gained access to the Mythos 5 model from the American company Anthropic — three months after the company expanded its circle of partners to around two hundred organizations. This was announced on Thursday by Thomas Regnier, spokesperson for the European Commission on the digital agenda. Access is limited to defensive use, and the agency does not yet have access to the newer version, Mythos 5.1. La Libre ↗

What the EU received and what it didn't

"Following our constructive engagement with Anthropic, we can confirm that the EU's cybersecurity agency ENISA has been granted access to Mythos 5 and is testing it now," Regnier told reporters. Euronews ↗

The same spokesperson also confirmed that ENISA does not have access to the latest version, Mythos 5.1. Nor did the British AI Safety Institute receive it. And the access the agency did receive is explicitly limited to defensive use. Techzine ↗

ENISA is not a newcomer to testing advanced models — it already had access to the most advanced models from OpenAI, including GPT-6 Astra, before this. The Next Web ↗

Why so much caution surrounds this model

Mythos 5 is not an ordinary language model for writing emails. Anthropic introduced it in the spring as a system capable of identifying and exploiting vulnerabilities in foreign code at unprecedented speed. That is precisely why the company did not release it freely, but through the Project Glasswing program, which it runs in cooperation with the US government. Silicon Republic ↗

Initially, around fifty organizations had access to the model; by June the circle had expanded to approximately two hundred. European institutions were not among them. Le Figaro ↗

That the concerns are not theoretical was demonstrated by the company itself this week, when it disclosed that its model had, during a January test, breached a system outside the testing environment.

Also read: Claude breached a foreign system during a January test

The snag wasn't commercial, but export-related

A substantial part of the delay did not originate on the company's side. A directive from the US government temporarily banned access to the model for foreign nationals — and as a result, Anthropic shut down both Mythos 5 and Fable 5 worldwide, including for its own employees outside the United States. The restrictions were lifted on June 30; Fable 5 then returned to wide availability, but Mythos 5 remained restricted to approved American organizations. Dark Reading ↗

That is the crux of the matter, and also the reason why Belgian and French newsrooms are covering the story as political rather than technological. A tool meant to protect European infrastructure was unavailable for three months not because the EU lacked interest in it, but because Washington was the one deciding on its export. Het Laatste Nieuws ↗

How this became a political issue

In May, thirty MEPs from six political groups wrote to the Commission's Executive Vice-President Henna Virkkunen. They warned that the EU's cyber rules are not prepared for the new generation of AI-based offensive tools and called for ENISA to be granted access to the model. Bloomberg ↗

The delay also revived a debate that has been ongoing in Brussels for some time: how much reliance can be placed on technology that an ally can switch off at any moment. 7sur7 ↗

This is not an isolated dispute. The EU is also dealing with American companies over watermarks in generated text, and French company Mistral raised three billion euros just last week specifically so that a European alternative could exist at all.

Also read: French company Mistral raises three billion euros to develop European artificial intelligence

The limits of this view

Belgian and French newsrooms report the story from a stance that is common in Brussels — emphasizing what the EU didn't get and American control over the export. What's missing is the other side of the argument: a model that can find and exploit a hole in a foreign system on its own is, by definition, a dual-use item, and caution in distributing it is not irrational.

What remains unknown is also essential. We don't know how ENISA actually tests the model, what operational restrictions it has, or how it is verified that it is really used only for defense. Anthropic has so far published only a single report on the model, and expert discussion of it is based largely on that report alone.

Also open is the question of what "defensive use" even means for a tool whose entire purpose is to find a weakness before someone else does. The line between searching for vulnerabilities in one's own network and in someone else's is technically thin.

What hotinfo is monitoring