Europe is now testing Mythos 5 and GPT-6 Astra for cybersecurity risks

The European Union's cybersecurity agency has gained access to two of the world's most advanced AI models for cybersecurity testing.

ENISA is now evaluating Anthropic's Mythos 5 and OpenAI's GPT-6 Astra, according to the European Commission.

The development follows months of discussion over whether European safety and security institutions would be able to independently evaluate increasingly restricted frontier AI systems.

Why these models matter

Advanced AI models are becoming increasingly capable at cybersecurity tasks.

That includes defensive work such as inspecting code and identifying vulnerabilities.

The same capabilities can also be used to search for weaknesses, develop exploits and automate parts of offensive cybersecurity work.

That dual-use nature has caused AI companies to place tighter restrictions around access to their most capable cyber models.

Anthropic in particular has limited access to some Mythos models to vetted organizations.

Europe wanted independent access

European officials have argued that governments need direct access to frontier systems if they are expected to evaluate their risks independently.

Earlier in September, European Commission officials said discussions with Anthropic and OpenAI about model access were still ongoing.

ENISA has now secured access to Mythos 5 and GPT-6 Astra and is testing the models.

The evaluations are expected to focus on cybersecurity capabilities and the risks associated with increasingly autonomous AI systems.

Frontier models are becoming strategic technology

The debate illustrates how quickly access to advanced AI has become a geopolitical issue.

Governments already treat leading-edge chips, defense technology and some cybersecurity tools as strategically sensitive.

Frontier AI models are beginning to receive similar treatment.

AI companies want researchers and governments to help evaluate their systems.

At the same time, giving powerful models to more organizations can itself introduce security concerns.

ENISA gaining access therefore represents more than another benchmark exercise.

It is part of a larger negotiation over who gets to independently inspect the most capable AI systems before and after they are deployed.