Months after its launch in April, Anthropic has finally granted access to Mythos, its artificial intelligence model specialised in cybersecurity, to the EU Agency for Cybersecurity (ENISA), a European Commission spokesperson confirmed to Bloomberg. Mythos was first presented in April and is considered sufficiently advanced at identifying computer vulnerabilities for the company to have limited access to a restricted group of pre-assessed institutions.
Around 40 entities are participating in the initiative called the Glasswing Project, including companies such as Apple, Amazon Web Services (AWS), Google, Nvidia and Microsoft. Access to the AI model has been heavily controlled due to its capabilities and potential impact on undesired access to sensitive data. The European Commission began negotiating access with Anthropic at the end of May, after Brussels pressured the company, but the talks triggered months of disputes over the type and scope of access.
Access may also have been conditioned by the restrictions imposed by the Trump Administration on foreign access to the Fable and Mythos models, which have since been eased. Although it guaranteed the EU that it could test the model — five months after its launch — Anthropic refused to submit its most recent model to the United Kingdom's AI Security Institute (AISI) for pre-launch testing.
The exclusion caused concern within the British Government, given fears that US technology companies may be adopting an increasingly restrictive stance towards foreign institutions. For Europe, the issue is no longer merely technological, as less access to the most advanced models also means less capacity to evaluate them and anticipate the risks associated with their use.
