Anthropic Mythos pilot opens to EU cybersecurity agency ENISA

Mythos is one of the most controversial model families in the frontier stack: a system explicitly designed to find and exploit software vulnerabilities autonomously. Putting it into the hands of ENISA, even under a tightly controlled pilot, is a watershed moment for how democratic governments intend to use offensive‑capable AI tools for defence. The Project Glasswing structure — selective access for vetted organisations, defensive‑only clauses, and pre‑deployment testing — mirrors OpenAI’s Trusted Access for Cyber program and signals an emerging norm: high‑end cyber models will not be released openly, but distributed under gated regimes to states and critical‑infrastructure operators.([techzine.eu](https://www.techzine.eu/news/security/141744/eu-cybersecurity-agency-gains-access-to-anthropic-mythos/?utm_source=openai))

For the AGI race, this is further evidence that red‑teaming and cyber‑security are becoming core monetization paths for frontier labs. Anthropic is building a coalition that includes US agencies, major banks and now an EU cyber authority, effectively turning Mythos into a de facto standard for AI‑assisted vulnerability discovery. That gives the company privileged feedback loops on how such systems behave in real networks — an information advantage that could transfer back into safer, more robust general‑purpose models. It also raises an uncomfortable prospect: as more powerful exploit‑writing systems proliferate among allied governments, the distinction between "defensive" and "offensive" AI capabilities becomes thinner and harder to police.