Anthropic Grants EU Cybersecurity Agency Access to Test Restricted Mythos 5 Model
Following months of regulatory negotiations, Anthropic has officially granted the European Union Agency for Cybersecurity (ENISA) access to test its high-performance frontier AI model, Mythos 5. European Commission spokesperson Thomas Regnier confirmed that ENISA is currently evaluating the model's cyber capability and security risks. This deal marks a major milestone in international AI governance and safety evaluations for high-risk frontier models. It highlights the complex interplay between US national security export controls, global regulatory demands, and cross-border cooperation in defending critical infrastructure from AI-assisted cyberthreats. The access agreement is only partial, as ENISA was granted testing rights to Mythos 5 but denied access to the newer Mythos 5.1 iteration. Negotiations were significantly delayed by White House access restrictions on foreign bodies, which also prevented international allies like the UK AI Safety Institute from accessing the latest Mythos version.
## BACKGROUND
Anthropic's Mythos series consists of highly capable language models designed with potent cybersecurity vulnerability discovery features, leading Anthropic to withhold general public release to prevent abuse by bad actors. Through Project Glasswing, Anthropic selectively shares Mythos with vetted organizations to preemptively find and patch security bugs before malicious hackers can exploit them.