David Sachs shared the details of the situation with Anthropic based on conversations with people inside and outside the US government.

Anthropic this week released models of the Mythos class under the trade name Fable. Fable is Mythos with guardrails. If the guardrails fail, Mythos' advanced cyber capabilities will be available to those who shouldn't. Anthropic itself widely promoted the idea that Mythos is a cyberweapon that needs to be regulated by the state, and advocated guardrails in Fable. In case of vulnerability, their responsibility will be patched.

A reliable partner who tested Fable found jailbreak guardrails. The administration asked Dario to fix the problem or remove the model. Dario refused.

In the blog, Anthropic stated that the jailbreak is not serious. But the partner and USG do not agree with this - such minimization does not correspond to their brand of AI security company.

Previously, Anthropic has always said that safety is the number one priority and should be taken extremely seriously. Here, the company has put a continuation of the sale of the consumer model vyshe besoposnosti.

In response, the Administration introduced export control (did it reluctantly). It is hoped that Anthropic will fix the problem, the control will be removed and Fable will be returned to public access as soon as possible. The administration is surprised by the company's refusal to comply with security requests, which it itself previously called its highest priority.

This is unrelated to previous DoW and Anthropic issues. The administration appreciates the technical capabilities of Anthropic and believes that the problem, although serious, can be easily solved. The ball is in Anthropic's court.