Anthropic's Claude Mythos model needed only hours to expose security flaws in highly sensitive U.S. government computer systems during a classified intelligence test, an official said.
Key Points:
- Anthropic's Mythos AI identified vulnerabilities in classified U.S. government systems within hours during an intelligence test, an official said.
- Finding a weakness is not the same as exploiting it, the official cautioned.
- More than 100 cybersecurity experts want the government to reverse export curbs that pulled Mythos and Fable 5 offline.
Mythos Test Exposes Classified Government Flaws
A U.S. official, speaking anonymously about the sensitive work, said intelligence agencies ran the exercise with the company under a program called Project Glasswing. The initiative gathered tech giants and other firms to harden critical software before flaws could cause severe harm to public safety and the economy. Anthropic had withheld a public release of Mythos, instead granting early access to a select group of companies so they could find and patch critical bugs ahead of any attacker.
Democratic Sen. Mark Warner of Virginia first referenced the tests on Jun. 11, at a hearing before the Senate Committee on Banking, Housing and Urban Affairs. He said the tool broke into almost all of the government's classified systems in hours rather than weeks, attributing the account to Joshua Rudd, who leads the National Security Agency and U.S. Cyber Command.
Identifying a weakness, the official cautioned, is not the same as exploiting it.
Also Read: Is The Anthropic Perp Sell-Off A Warning For Pre-IPO Crypto Bets?
Cybersecurity Experts Challenge Export Curbs
More than 100 cybersecurity leaders, including figures from Adobe and Nvidia, have pressed the government to reverse the export curbs and commit to a transparent process for weighing AI risk. They called Mythos capable at finding and weaponizing software flaws, yet not uniquely so, since many of them lean on rival and open-source models for the same work. China's systems trail the best American models by only months, they argued, making the timing especially risky.
The pushback follows a Jun. 12 export directive that barred foreign nationals from Mythos 5 and Fable 5, the broader release of the Mythos tier. Anthropic disabled both models for every customer to comply, while maintaining it saw no security basis for the government's move.
The directive followed an executive order from President Donald Trump that set up federal vetting of the most advanced AI systems for up to a month before release. Participation by developers would be voluntary, the order said, though tensions with the safety-focused company have run high.
Mythos Cyber Record Predates The Clash
The model had arrived this spring carrying a documented knack for surfacing software flaws that even seasoned human researchers miss. Britain's AI Security Institute earlier cleared Mythos on expert-level capture-the-flag challenges that no system had solved before, with the model passing 73% of them. Mozilla separately credited an early version with finding 271 vulnerabilities in Firefox, which it patched in the browser's 150th release.
Read Next: Mane City Mobile Lands On iOS And Android In 100+ Countries