Anthropic's Fable Is Locked Down As US Takes AI Safety Into Its Hands

But the company's resistance may be overshadowed by a larger question: what happens when a founder-led AI company's own warnings about existential risk get taken literally by policymakers?

Anthropic Advocated Its Way Into Regulation With Fable

This is what happens when you advocate loudly for AI safety regulation. The government listens.

Anthropic built its entire corporate narrative on the premise that models like Mythos pose such severe risks that they should never be released to the public. The company stated plainly in April that Mythos was "too dangerous" for broad release because "the fallout for economies, public safety, and national security could be severe." Anthropic launched Fable 5 as a compromise, a heavily guardrailed alternative designed to capture some of Mythos' capabilities while containing its worst behaviors.

The contradiction is striking. Anthropic spent this spring fighting a Pentagon supply-chain-risk designation that excluded it from Defense Department contracts, a label a federal judge suggested looked like punishment after the company refused to let the military use Claude for mass surveillance and autonomous weapons. Now the same government that tried to push Anthropic out is invoking national security to lock its models away from foreign users. The company is treated as both a liability to do business with and a capability too potent to share.

But guardrails are not walls.

No AI model in history has avoided jailbreaking entirely. Anthropic invested tremendous effort in red-teaming Fable 5, searching for exploitable weaknesses in its defenses. The company even leads the industry in "observability research," the effort to understand how large language models actually behave at scale. And yet, as Anthropic itself knows, no one truly understands these models completely. Their behavior remains partly opaque even to their creators.

If Mythos truly represents the danger Anthropic claimed, then Fable 5 likely carries some version of that same risk. The guardrails reduce the probability of harmful outputs, but they do not eliminate it.

The Commerce Department directive, issued by Secretary Howard Lutnick under the Bureau of Industry and Security, represents the latest escalation in government AI regulation. This specific action follows a broader White House approach to AI regulation that Sean Cairncross has helped shape.

Crucially, this is a mandatory licensing action, not the voluntary testing framework the administration rolled out earlier this month. Per Axios, the executive order on pre-deployment testing is explicitly voluntary and avoids a licensing regime, a structure White House chief AI adviser David Sacks secured to prevent what he calls "regulatory capture" of the largest labs. The export control is a separate, binding measure.

Per Commerce's letter, a license is now required for any export, re-export, or domestic transfer of the models, Anthropic must file additional applications for individually validated licenses, and failure to comply carries financial and civil penalties