BitGo CEO Mike Belshe is pushing back against a viral claim that Anthropic's Mythos AI model broke into nearly all NSA classified systems. The assertion, which Senator Mark Warner cited from NSA Director General Joshua Rudd, said Mythos 'broke into almost all of our classified systems... in hours.' But Belshe and others say the supposed intrusion was actually an authorized red-team test on the NSA's own networks — not an unauthorized breach.
How the claim caught fire
Warner's statement, based on Rudd's remarks, spread quickly. The Economist's Shashank Joshi cautioned that the quote shouldn't be read literally. It depended on Mythos working with other tools under specific conditions, he said. Still, the idea of an AI model cracking the most secure government systems rattled observers.
What Mythos really did
The US government was already a partner of Anthropic. Mythos had been deployed to government cyber defenders through Project Glasswing since April. The red-team exercise gave Mythos access to NSA networks — it wasn't a break-in. Belshe's point: calling it a breach misrepresents what happened. Anthropic describes Mythos as the strongest cyber model in the world, but its capabilities were tested in a controlled setting.
Why the model was disabled
On June 12, Anthropic disabled both Mythos models. The reason wasn't a security incident. It was a US export-control directive. Anthropic objected to recalling a model used by hundreds of millions over one narrow flaw. AI researcher Pedro Domingos argued export controls were responsible, given the model's powerful hacking abilities — not a breach.
What happens next
Anthropic is working to restore access to Mythos and drafting a shared risk framework with the White House. Whether the export controls will remain in place or be adjusted as those talks continue is unclear.
