Anthropic released Fable, a safety-guarded version of its supposedly dangerous Mythos model, only to face immediate export restrictions from the US government. Officials cited national security concerns over potential jailbreak vulnerabilities, forcing the company to disable access for all customers including foreign nationals. The timing raises questions about whether the cautious rollout strategy actually worked.

The conflict centers on a jailbreak technique that reportedly exposed minor vulnerabilities already discoverable by other models. Amazon, an Anthropic investor and inference provider, apparently reported the bypass method. Senior Anthropic staff are now in Washington D.C. disputing what they call a misunderstanding, while the government frames it as leadership negligence regarding legitimate security risks.

This incident reflects broader economic tensions between frontier AI labs and software companies. Anthropic and OpenAI have burned tens of billions building models that open source alternatives commoditize. Meanwhile, Satya Nadella argues companies should build learning loops combining human capital with AI capabilities rather than relying on interchangeable models.

The real battle isn't just technical safety—it's about who controls the user touchpoint in the AI value chain. Nadella warned against a future where only a few models accrue all value while hollowing out industries, echoing concerns about globalization's displacement effects. The political economy won't tolerate extreme concentration of AI power.