Mythos Allegedly Breaches Almost All NSA Classified Systems
The Economist, as cited by TechTimes, reports that Senator Mark Warner relayed that Gen. Joshua Rudd told him Anthropic's Mythos, in a June 11 authorized red-team exercise, "broke into almost all of our classified systems, not in weeks, but in hours." The Economist editor Shashank Joshi, who wrote the original quote, later clarified the claim "should not be read literally," noting it depended on Mythos working alongside other tools under specific conditions. On June 12, the U.S. Commerce Department issued an export-control directive barring all foreign nationals including Anthropic's own non-citizen employees from accessing Fable 5 and Mythos 5. Anthropic, unable to selectively comply, suspended both models for all users worldwide, per Fortune. Anthropic called the directive "a misunderstanding" and says the cited trigger was a narrow jailbreak - asking the model to read a codebase and fix flaws - that also applies to rival models including OpenAI's GPT-5.5. BeInCrypto and Digg document expert pushback; independent forensic verification of the red-team scope remains outstanding.
What happened
The Economist first published the account, citing Senator Mark Warner (vice chair, Senate Intelligence Committee), who said NSA director Gen. Joshua Rudd told him Anthropic's Mythos ("Mythos 5") "broke into almost all of our classified systems, not in weeks, but in hours" during an authorized red-team evaluation on June 11. The Economist editor Shashank Joshi, who wrote the original passage, later stated on X that his report had been widely misread - the claim "should not be read literally," as it depended on Mythos working in conjunction with other tools under particular conditions, per BeInCrypto's June 21 reporting of Joshi's clarification.
Government action
On June 12 at 5:21 PM ET, Anthropic received a Commerce Department export-control directive barring all foreign nationals - including Anthropic's own non-citizen employees - from accessing Fable 5 and Mythos 5, per Fortune and Anthropic's own statement. Because selective compliance would have required blocking part of its own staff, Anthropic suspended both models for all users globally. Access to less powerful models, including Claude Opus 4.8, was not affected, per Fortune. CybersecurityNews notes this marks the first time the United States has applied export controls directly to an AI model rather than to hardware or chips - a landmark regulatory precedent in AI national security governance. Allied Five Eyes governments were reportedly caught off guard, with permissions revoked for agencies, banks, and major firms without prior warning, per CybersecurityNews.
Anthropic's dispute
Anthropic states the directive did not provide specific details of the government's national security concern. Officials told Anthropic the decision was triggered by a technique to bypass Fable 5 safeguards - a narrow jailbreak consisting of asking the model to read a codebase and fix identified flaws, per Fortune. Anthropic says the same jailbreak elicits similar behavior from OpenAI's GPT-5.5, and argues that "if this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers." Anthropic called the government's position "a misunderstanding" and said it is working to restore access and drafting a shared risk framework with the White House (Fortune, BeInCrypto).
Earlier BBC investigation
The BBC reported in April that Anthropic was "investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments," and at that time had no evidence its own systems were affected. BBC coverage framed that earlier incident as likely involving misuse of legitimate access rather than classical intrusion.
Expert dispute and verification
BitGo CEO Mike Belshe publicly rejected the viral breach claim as false (BeInCrypto). Digg aggregated expert skepticism that observed red-team behavior reflected autonomous model capability rather than exploitation of existing software vulnerabilities. Analyst Kyle Chase noted the red-team result was a controlled test, and a separate narrow jailbreak flagged by Amazon was cited as the government's direct trigger, per BeInCrypto. Independent forensic corroboration of the full red-team scope remains outstanding.
Implications for practitioners
For security and governance teams, the episode highlights three separable risk vectors:
- •model-in-the-loop capability to discover or weaponize software vulnerabilities in a directed red-team setting
- •export-control governance now applied at the AI model level rather than at the hardware level for the first time
- •the operational risk of third-party vendor access controls. The policy response illustrates that contested red-team claims can trigger near-immediate access restrictions at commercial scale
Scoring Rationale
The episode combines a Senate-level national security claim about a frontier AI model's red-team performance with the first-ever U.S. export control applied directly to an AI model rather than hardware - a landmark governance precedent affecting hundreds of millions of users. The primary source editor's own clarification that the breach claim 'should not be read literally' adds a contested-framing dimension that practitioners must navigate. Together, the model-level export control precedent and the policy response speed elevate this above a standard security incident.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
