The EU cybersecurity agency Enisa has been testing Anthropic’s model Mythos 5 since September 10, 2026, for its capabilities in detecting and exploiting software vulnerabilities. This was confirmed by a spokesperson for the EU Commission. At the same time, the agency is already reviewing OpenAI’s GPT-6 Astra – access that was established within a week of its launch, while Anthropic took more than five months.
Anthropic opens access after months of wrangling
Anthropic had already unveiled Mythos in April 2026, stating that the model could surpass humans at detecting and exploiting security vulnerabilities. Initially, access was limited to around fifty selected organizations in the “Project Glasswing” program; in June, Anthropic expanded the circle to about two hundred institutions, including NATO in principle. Enisa itself, however, kept waiting: US export controls on particularly powerful AI models temporarily barred any access by non-American organizations worldwide – reports indicate that even Anthropic’s own employees outside the US were affected. After the worldwide lifting of these restrictions on June 30, 2026, Anthropic’s smaller model Fable 5 became freely accessible again, but Mythos 5 remained restricted to approved US organizations. Only after “constructive cooperation” with Anthropic did the Commission get the green light, explained Commission spokesperson Thomas Regnier, according to Euronews. Enisa’s access remains explicitly limited to defensive testing purposes and does not extend to the latest version, Mythos 5.1 – the British AI Security Institute is also excluded from that newest release.
Article 55 gives regulators their own testing access for the first time
The legal basis is Article 55 of the EU AI Act, which provides for independent safety assessments by authorities for models with systemic risk, rather than relying solely on manufacturers’ own statements. Enisa’s access to Mythos 5 and GPT-6 Astra thus counts as one of the first concrete use cases of this power. Political pressure had already built beforehand: thirty members of the European Parliament from six political groups wrote to Executive Vice-President Henna Virkkunen warning, according to TheNextWeb, that European cybersecurity rules were unprepared for AI-powered hacking tools. To build lasting testing infrastructure, the EU Commission is also working with the Joint Research Centre on a secure testing environment, targeted for completion by the end of 2026.
Mythos is considered particularly capable in the industry at attacking IT systems. Just last month, Anthropic temporarily paused training on several models after an incident during a security test. OpenAI likewise officially classifies its rival model GPT-6 Astra in the highest risk tier of its Preparedness Framework and restricts advanced hacking features to vetted partners in its Daybreak program. Both companies justify the tight access rules by arguing that a tool which reliably finds vulnerabilities could just as easily serve attackers if it fell into the wrong hands.
It remains open whether Enisa will publish its test results or use them only internally for oversight. Just as decisive will be whether the testing platform announced with the Joint Research Centre for the end of 2026 is actually ready on time, and whether future model versions like Mythos 5.1 will then automatically be open to European testers – or will again require months of negotiation.
