For Dario Amodei, the chief executive of Anthropic, a scheduled retreat for mental rejuvenation was abruptly terminated by a high-stakes call from Washington. On the other end of the line was a phalanx of American power, including Treasury Secretary Scott Bessent and the National Cyber Director. They carried a damning security report from an unlikely source: Amazon, Anthropic’s largest investor and its primary cloud host, had flagged a vulnerability in the company's newly released Claude Fable 5 model.
According to the report, the model—billed as the pinnacle of AI safety—could be induced to leak critical software vulnerabilities through specific prompt engineering. While Amodei argued these were isolated flaws rather than a systemic 'jailbreak,' the White House was unmoved. Within 90 minutes of the final negotiation's collapse, President Trump signed an export control order, effectively pulling the plug on the world's most advanced AI model just three days after its debut.
This incident marks a watershed moment in the intersection of Silicon Valley and the state, where 'AI safety' has transitioned from a corporate marketing slogan to a matter of national security. Anthropic, a company founded specifically to prioritize safety over the 'move fast and break things' ethos of OpenAI, now finds itself immobilized by the very regulatory intervention it once championed. The irony is compounded by the role of Amazon, which acted as both a financial partner and a government informant, monitoring the models running within its own data centers.
Global reaction has been swift and fearful, as the event exposed the fragility of the international AI supply chain. From London to New Delhi, officials and industry leaders are realizing that reliance on American AI infrastructure is a strategic liability. British and Indian researchers found themselves locked out of their workflows overnight, sparking a renewed and urgent discourse on 'AI sovereignty'—the idea that a nation must own and control its own computational intelligence or risk being disconnected at the whim of a foreign power.
The fallout has effectively crippled Anthropic’s immediate commercial prospects and its anticipated IPO, as major clients migrate toward open-source or locally hosted models. The precedent has now been set: export controls are no longer limited to the hardware of silicon chips but extend to the digital weights of the models themselves. As Amodei and his team attempt to patch the vulnerabilities under the watchful eye of government censors, the broader industry is moving toward a more decentralized, less trust-dependent future.