Lessons from Anthropic’s failed Fable 5 release

Earlier this month, Anthropic launched Claude Fable 5, an AI model the company built after cooperating with the federal government in good faith. Anthropic had allowed government reviewers access to its underlying model of Mythos, took their feedback seriously, and added guardrails against high-risk uses in cybersecurity and biological materials. Within days of Fable’s public release, the Department of Commerce hit it with export control restrictions, reportedly after a phone call from Amazon CEO Andy Jassy to the White House.  

Anthropic did everything right. It got punished anyway. 

Days before the launch of Fable, the Trump administration issued its Executive Order encouraging artificial intelligence companies to voluntarily give the federal government up to 30 days of access to their “covered frontier models,” to promote security and protect critical infrastructure. But “covered frontier model” remains undefined.  

The order gives a litany of departments 60 days to set the threshold, including the secretary of the Treasury; the secretary of Defense (via the National Security Agency); the secretary of Homeland Security (via the Cybersecurity and Infrastructure Security Agency, currently without a director); the assistant to the president for Science and Technology Policy; and the secretary of Commerce (via the National Institute of Standards and Technology). Before this process can evaluate a single model, it first has to be built. This will take time. 

Other than the scheduling issues such a deadline imposes, the learning curve for review is too high. The current executive order directs that the Treasury secretary lead the process despite having little institutional knowledge........

© The Hill