MythosWatch is building a public record of who can access Anthropic's restricted Claude Mythos Preview model. The results so far: 12 named partners, 5 entities in negotiations (including the NSA, UK AI Security Institute, and multiple US departments), 2 blocked entities, and over 40 organizations with confirmed but undisclosed access. The site aggregates 39 public signals to map what it calls, somewhat grandly, "the most powerful AI."
Where Mythos goes isn't random. Distribution tracks closely with the Five Eyes intelligence alliance. The US, UK, and Australia have named participants or reported government use. China is actively blocked under US export controls. The Pentagon is legally barred from access while the NSA reportedly has access. Same government, different rules. The European Union has zero named entries but shows regulatory activity through financial and cybersecurity authorities. Deutsche Bundesbank has already called for institutional access.
A security incident makes the tracking urgent. Bloomberg reported unauthorized users accessed Mythos through a third-party vendor environment, sharing screenshots and a live demo via a private Discord channel. Anthropic confirmed an investigation but said core systems weren't impacted.
The Hacker News community pushed back on MythosWatch's methodology, suggesting the data might contain hallucinations. The tracker links every entry to sources, though, and uses clear labels: "Official" means a primary source names the entity, "Reported" means credible reporting exists without direct confirmation. The data is CC BY 4.0 licensed. Whether the methodology holds up under scrutiny matters less than the fact that someone is trying to build this record at all.