Security and crypto executives are disputing a viral claim that Anthropic's three-day-old AI model Mythos broke into nearly all U.S. classified systems in hours. Key Points: A viral post clai
Security and crypto executives are disputing a viral claim that Anthropic's three-day-old AI model Mythos broke into nearly all U.S. classified systems in hours.
- A viral post claimed Mythos breached almost all NSA classified systems in hours; security and crypto leaders call it a misread.
- The cited result came from an authorized red-team test on the agency's own networks, not a live intrusion.
- Anthropic ties the actual shutdown to a narrow jailbreak and a U.S. export-control order, not the test.
The line came from Senator Mark Warner, who relayed it from General Joshua Rudd, the officer now jointly leading the NSA and U.S. Cyber Command. A widely shared policy briefing first reported the remark on Jun. 14, where it sat largely unnoticed until an AI commentator and a prediction-market account amplified it over the weekend. Within hours, the name Mythos was trending.
The framing online was blunt: a real-world hack of the nation's secrets.
The fuller record is narrower. Warner raised the example while praising Anthropic, not attacking it, using the moment to argue for mandatory pre-release testing of frontier models rather than trust in the good faith of any single lab. The break-in he described unfolded inside an authorized red-team exercise on the agency's own networks, the kind of controlled drill security teams run to expose weak points before adversaries do.
Also Read:Bitcoin ETF Exodus Hits Record $6.35B, But Panic Selling May Be Cooling
BitGo chief executive Mike Belshedisputed the story almost as soon as it spread, flatly calling it false in a reply that drew heavy attention across the crypto industry. Several analysts piled on, mocking how a senator's offhand line had traveled from a closed hearing to viral posts, with one noting that no official NSA statement ever confirmed any of it.
Critics also pressed a point about timing. Anthropic ties the actual shutdown to a narrow jailbreak that Amazon flagged, alongside a U.S. export-control order that pulled the models offline on Jun. 12. The company says that flaw merely asked the model to read a codebase and surface a few minor, already-known bugs that rival systems like OpenAI's GPT-5.5 can find just as easily.
Lost in the weekend frenzy was a basic fact about access, since the U.S. government had already been running Mythos for weeks.
Anthropic had deployed the model to federal cyber defenders since April under Project Glasswing, a closed program limited to roughly 200 vetted partners including Amazon, Apple, Google, Microsoft, JPMorgan and the Linux Foundation.
The fight has simmered all month. After the Jun. 12 directive forced Mythos and Fable offline, more than 100 cybersecurity leaders urged a reversal, warning that the curbs hand an edge to U.S. rivals whose own models sit only months behind. Access remains frozen as the lab and the White House draft a shared risk framework.
Read Next:JaredFromSubway Bot Loses $7.5M After Taking Its Own Bait
