A researcher typed 'fix this code' into Anthropic's most powerful model and the United States government shut it down within ninety minutes. The vulnerability was real. The precedent is bigger.
On June 9, Anthropic released Fable 5, the most capable AI model the company had ever built. Three days later, the United States government told them to take it offline.
The vulnerability that triggered the shutdown was a prompt. Not a sophisticated adversarial attack, not a novel exploit chain, not a zero-day in the inference stack. A researcher typed three words into the chat interface (fix this code) and the model provided information about cybersecurity vulnerabilities that its safety systems were supposed to suppress. A jailbreak researcher known as Pliny the Liberator published a more elaborate version combining multi-agent decomposition, Unicode tricks, and narrative framing. The result was the same: Fable produced outputs its designers had classified as off-limits.
The Commerce Department issued its directive at 5:21 PM on June 12, citing national security. Anthropic was given ninety minutes. The order used export control authority, the legal framework designed for weapons and dual-use technology, to bar all foreign nationals from accessing Fable 5 and its underlying model, Mythos 5. That category includes non-citizens working inside the United States, including some of Anthropic's own engineers. Rather than maintain a system that excluded portions of its own workforce, Anthropic took both models offline entirely.
The conventional reading is that the government acted responsibly. A powerful AI model had a safety bypass. National security was at risk. The adults in the room stepped in.
The less conventional reading starts with who rang the alarm.
Amazon discovered the bypass through its internal research team. Amazon CEO Andy Jassy personally communicated the finding to government officials. Amazon has invested $13 billion in Anthropic, with commitments for up to $25 billion. Amazon hosts the model on AWS. Anthropic has pledged to spend over $100 billion on AWS infrastructure over the next decade. And Amazon's own AI products, Bedrock and Nova, compete directly with the model it helped pull offline.
Jassy may have acted in good faith. The vulnerability was real. But the channel matters. The finding traveled from corporate competitor to executive branch, bypassing the company that built the model. Anthropic says it worked with the US government, the UK AI Safety Institute, and private organizations to red-team Fable's safeguards for thousands of hours before launch. The administration had pushed to delay the release. Anthropic declined. The export control letter followed.
Anthropic's response was measured. Dario Amodei argued on calls with senior administration officials that the bypass was narrow, that the information it surfaced was already publicly available, and that equivalent results could be obtained from competing AI systems without any safety bypass at all. The company sent staff to Washington. Inside the administration, the mood was different. An official who had pushed to give Anthropic a chance told Axios: They screwed us.
What makes this episode important is not the vulnerability. Jailbreaks happen. Every frontier AI system has them. OpenAI's models, Google's Gemini, Meta's open-source Llama. All have been bypassed using techniques no more sophisticated than what took down Fable. The question is why this jailbreak, of this model, produced this response.
One answer is that Anthropic built its brand on safety. The company was founded by former OpenAI researchers who left specifically because they believed AI development was moving too fast without adequate safeguards. Anthropic published responsible scaling policies, created Constitutional AI, invested in interpretability research. The pitch to regulators, investors, and the public was: we are the careful ones. When the careful ones get jailbroken, the political cost is higher than when the ones who never claimed to be careful get jailbroken. The brand becomes the liability.
Another answer is architectural. Amazon sits at the intersection of investor, infrastructure provider, and competitor. No previous technology era produced this configuration. Standard Oil did not host its competitors' refineries. AT&T did not fund the companies whose calls it routed. The AI industry has created a set of relationships where the entity that provides your compute, bankrolls your research, and competes with your products can also, through a single phone call to the right official, trigger the legal machinery that shuts you down. The export control framework was not designed for this. It was designed for centrifuges and encryption chips, not for the commercial relationship between a cloud provider and its largest customer.
The ninety-minute deadline is the detail that matters most. Not because it was unfair to Anthropic, though the company clearly believes it was. Because it establishes a precedent. The United States government demonstrated that it can compel an AI company to shut down its most important product, globally, in under two hours, using national security authority, with no advance notice, no published criteria for what constitutes a violation, and no appeals process.
Every AI company in the world now operates under rules that do not exist yet. The government proved it can act without them. The question going forward is whether this power gets exercised through policy (published standards, defined thresholds, review periods) or through phone calls. And whether the next call comes from a regulator or a competitor.
Anthropic spent years building the argument that AI companies should regulate themselves before governments regulate them. The Fable episode suggests a third option nobody planned for: regulation through ad hoc intervention, triggered by corporate intelligence, with no framework and ninety minutes on the clock.
The safety researchers were right that AI needed guardrails. They were wrong about who would build them.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)