BitGo CEO Calls "BS" on Mythos NSA Hack Tale: Red-Team Test, Not a Breach 🛡️

BitGo CEO Mike Belshe has rejected a viral claim that Anthropic's Mythos model breached nearly all of the National Security Agency's classified systems, calling the story false as it spread across X this weekend. His pushback targets posts that recast the government shutdown of a three-day-old model as a real-world hack. The claim originated with Senator Mark Warner, vice chair of the Senate Intelligence Committee, who cited NSA director General Joshua Rudd during a conversation with The Economist. "This tool broke into almost all of our classified systems, not in weeks, but in hours," Warner said, according to the outlet, while praising Anthropic and arguing for faster pre-release testing of frontier models.

The detail missing from viral recaps, according to subsequent clarifications, is that the exercise was an authorized red-team test on the agency's own networks rather than an outside intrusion. Shashank Joshi, the Economist editor who published the quote, cautioned on June 21, 2026 that the line "should not be read literally" and said it depended on Mythos working alongside other tools in particular conditions. Anthropic had deployed the model to government cyber defenders through Project Glasswing since April, and the US government was already a Mythos partner before the test was described.

Belshe, co-founder and chief executive of digital-asset custodian BitGo, responded to one of the circulating threads bluntly, writing "I'm calling BS on this." Other commentators pointed to the same gap, with analyst Kyle Chase noting the activity was a test and arguing that a separate jailbreak flagged by Amazon was the actual trigger for government action. Anthropic's own statement supports that reading, saying the flagged jailbreak simply asked the model to read a codebase and fix flaws. The technique surfaced a few minor, already-known bugs that rival models such as OpenAI's GPT-5.5 can also find.

The company disabled both models on June 12 to comply with a US export-control directive, not because of any operational breach, and objected to recalling a model used by hundreds of millions of people over one narrow flaw. Whether the test justified pulling the models remains contested, with AI researcher Pedro Domingos arguing that the export controls were responsible given the model's powerful hacking capabilities, while Anthropic itself calls Mythos the strongest cyber model in its lineup.