The U.S. government ordered Anthropic to immediately shut down worldwide access to Claude Fable 5 and Claude Mythos 5 on Friday, citing national security concerns. The directive, received at 5:21 pm ET, forces the company to disable both models globally—not just for foreign nationals the export control order nominally targets. Anthropic complied but published a lengthy blog post arguing the government got it wrong. Access to other Claude models remains unaffected. The action represents the first time peacetime federal authorities have compelled an AI lab to pull deployed models from commercial use, establishing a precedent that will shape how frontier labs calculate risk going forward.
Mythos is Anthropic's most capable model, previewed in April and kept restricted because of what the company described as exceptional ability to find security vulnerabilities. According to Anthropic, Mythos identified flaws in every major operating system and web browser it tested. Rather than release it broadly, the company launched Project Glasswing, sharing it with roughly 50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike for defensive cybersecurity work. Fable 5, released just three days before the shutdown, was a guardrailed version designed for general release. Benchmark tests from Vals AI rated it the most capable publicly available AI model at launch.
The Wall Street Journal reports that Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent and other officials that Amazon researchers used Fable 5 to obtain information that could be used in cyberattacks. The government subsequently imposed the export control ban. An Amazon spokesperson said governments seeking counsel on security risks is not uncommon but declined to share discussion details. The Information and Reuters corroborated that Amazon, a major Anthropic investor, communicated security concerns. The timing raises obvious questions about competitive dynamics—Amazon competes directly with Anthropic in the foundation model market through its own investments and AWS Bedrock platform.
The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused.
David Sacks, Trump's former AI czar who now co-chairs the President's Council of Advisors on Science and Technology, offered a different account. Sacks claimed a highly credible trusted partner of both Anthropic and the government came forward with a jailbreak. The administration asked CEO Dario Amodei to fix the jailbreak or de-deploy the model, and Amodei refused. Anthropic countered in its blog post that the capabilities causing government concern are already available in other publicly accessible models. The government's directive is framed as export control but Anthropic says its understanding is the underlying concern is a claimed Fable 5 jailbreak, for which the government has provided only verbal evidence of a potential narrow, non-universal exploit.
The irony is sharp: Anthropic's cautious approach may have triggered the crackdown. The company built a reputation on safety-first development, publishing detailed capability assessments and implementing staged releases. That transparency gave regulators a clear target. Cybersecurity researchers had already criticized Fable's guardrails as excessive—Valentina Palmiotti of IBM X-Force noted it rejects requests tangentially cyber-related, even innocuous tasks like reading a blog post. Matt Suiche observed that asking it to write secure code triggers downgrade to Claude Opus 4.8 because anything in the lexical field of cybersecurity trips guardrails. The model appeared to be keyword-based rather than contextually intelligent about risk.
The timing coincides with a broader regulatory squeeze on frontier AI labs. A coalition of state attorneys general opened an investigation into OpenAI on the same Friday, with New York's AG serving a subpoena seeking documents on advertising, user engagement, model sycophancy, handling of consumer and health data, and treatment of minors and seniors. Florida Attorney General James Uthmeier sued OpenAI and Sam Altman earlier this month, claiming they ignored safety warnings and put children at risk. OpenAI still faces lawsuits over copyright infringement and ChatGPT's alleged role in user suicides. The coordinated pressure suggests state-level enforcers see an opening to shape AI governance where federal frameworks remain nascent.
The geopolitical implications extend beyond U.S. borders. Anthropic's suspension came shortly after announcing a partnership with Tata Consultancy Services to expand enterprise AI adoption in India. The directive requires suspending access for all foreign nationals, including Anthropic's own foreign national employees. Indian founders, investors, and policy experts are debating whether to accelerate domestic AI capabilities, invest in open-source alternatives, or continue relying on U.S. frontier providers. For some, the episode is a wake-up call on technological dependence. For others, it underscores that access to critical AI systems can be shaped by geopolitical decisions beyond their control.
The incident crystallizes tensions between AI acceleration and control that have simmered since ChatGPT's release. Labs face a Catch-22: demonstrate capability to attract capital and customers, but revealing capability invites regulatory intervention. Safety warnings intended to show responsibility can be read as admissions of danger. The government now has a playbook for rapid model shutdown based on claimed security concerns, with export control authority providing legal cover for actions that affect all users globally. Whether this represents genuine risk mitigation or safety theater dressed in national security language will depend on facts the government has not yet disclosed publicly.
The competitive dynamics are impossible to ignore. Amazon funds Anthropic but competes with it. OpenAI faces antitrust scrutiny while Anthropic gets shut down. The Information reports the White House is unlikely to extend similar restrictions to other AI companies and is privately blaming Anthropic's handling of alleged jailbreak vulnerabilities. If one company's models are forced offline while competitors' comparable systems remain available, the action becomes a market intervention with winners and losers, not a horizontal security measure. The question is whether regulatory capture is occurring—and if so, whether it flows through traditional channels or through the informal counsel of well-positioned CEOs with competing interests.
For venture investors, the calculus shifts immediately. Capital has poured into frontier labs on the assumption that regulatory risk is manageable and that labs with strong safety postures face less exposure. This week flipped that logic. Labs that move carefully and document risks may be more vulnerable than those that ship fast and stay quiet. The decision to invest in closed versus open models, U.S.-domiciled versus internationally distributed labs, and companies with versus without deep-pocketed strategic investors now carries regulatory risk premium that was theoretical last week and concrete today. Meanwhile, Anthropic had filed confidentially for IPO on June 1 at a $965 billion valuation. How the shutdown affects that timeline and valuation will signal how investors price regulatory intervention risk going forward.