Anthropic Mythos 5 cleared for select US critical‑infra users

Mythos 5 coming back—albeit in a constrained way—is a reminder that frontier AI capability is now inseparable from national security policy. Two weeks ago, Washington’s abrupt order forcing Anthropic to cut off all foreign access to Mythos 5 and Fable 5 sent a chill through the ecosystem; now, a carefully negotiated partial reopening for US critical‑infrastructure operators sketches what a new normal might look like.

For the race to AGI, the message is that the most capable models will increasingly live behind layered trust boundaries: project‑based consortia like Project Glasswing, government‑approved customer lists, and export‑style controls over who can touch which weights. That favors players with deep government relationships and mature safety tooling, because only they can credibly argue for expanded access after a scare like Mythos reportedly finding vulnerabilities in classified systems within hours. It also underscores how intertwined Anthropic and OpenAI’s fates are becoming: restrictions on one firm’s cyber‑capable model now trigger compensating scrutiny of the other’s releases. Over time, this could push leading labs toward differentiated model lines—“civilian,” “defensive,” and “high‑risk research”—rather than a single general‑purpose frontier system.