Reporters this weekend walked back the viral claim that Anthropic's Mythos model breached classified National Security Agency (NSA) systems, but the bigger US story of what roughly 200 elite firms lost when the model went dark on 12 June stays buried.
The clarification came from Shashank Joshi, defence editor at The Economist. He confirmed on X on 21 June that he had accurately quoted Senator Mark Warner, vice chair of the Senate Intelligence Committee, but said the line had been stripped of crucial context as it tore through social media.
What Warner Actually Said
Warner, speaking on 11 June, relayed what General Joshua Rudd, the four-star officer who jointly leads the NSA and US Cyber Command, had told him in a private briefing. Mythos, Rudd said, 'broke into almost all of our classified systems, not in weeks, but in hours' during authorised testing. The exercise was a red team drill on government networks, not an outside intrusion, and Warner cited it to argue for faster pre-release testing of frontier artificial intelligence (AI) models, not to condemn Anthropic.
BitGo chief executive Mike Belshe rejected the viral 'NSA confirms' framing outright, calling it false. No official NSA statement exists, only a senator publicly recounting what a general told him.
The 200 Firms in Project Glasswing
While the debunking played out, the more material US story sat in the background. Mythos and its restricted sibling Fable 5 were both pulled by Anthropic on 12 June, after the Commerce Department, citing national security, ordered the company to cut off foreign national access. Anthropic complied by disabling commercial access and public systems within hours, though select domestic enterprise partners in Project Glasswing retained their preview pipelines.
Both versions powered Project Glasswing, the cybersecurity consortium Anthropic launched in April. Launch partners included Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. By late May, the consortium had grown to roughly 200 organisations across 15 countries, according to CyberScoop. These firms quietly used the unguarded version of the model, priced at $25 (£19) per million input tokens and $125 (£95) per million output tokens, to scan their own code for flaws.
Why US Customers Should Care
The shutdown changed the defensive posture of every name on that list overnight. Americans bank with JPMorgan, store data on Amazon and Apple, and use federal services that may have been screened by Mythos for cybersecurity flaws. Mythos Preview surfaced more than 10,000 high or critical severity vulnerabilities between April and late May, Anthropic reported. Cloudflare found 2,000 bugs across critical systems, 400 of them rated high or critical. Mozilla patched 271 vulnerabilities in Firefox 150 while testing the model.
Andy Jassy, Amazon's chief executive, raised the concerns that triggered the order with Treasury Secretary Scott Bessent, according to Fortune, after Amazon researchers coaxed a Mythos class model into describing restricted cyberattack methods.
White House Framework Talks
Anthropic is now drafting a shared risk framework with the White House, Politico reported on 18 June. The proposed standard would score how far a bypass went, what capabilities became reachable, and whether the exposed behaviour had real operational effects. President Donald Trump told Axios on 21 June that he no longer viewed Anthropic chief executive Dario Amodei as a national security threat after meeting him at the G7 summit, calling him 'nice' and 'smart.'
For standard commercial users and international teams cut off by the directive, access has now been offline for 10 days. Missing out on the elite pipeline, these security teams have shifted to Anthropic's Opus 4.8 as a fallback: a model the company itself does not call the strongest cyber defender it has built.
