White House Says Anthropic Refused to Fix AI Jailbreak Before Model Shutdown

The US administration has claimed that Anthropic refused to fix or withdraw its Fable 5 AI model after being warned of a jailbreak that exposed advanced cyber capabilities, escalating a public dispute over the government's unprecedented decision to disable two of the company's flagship models.

The allegations were made by David Sacks, co-chair of the President's Council of Advisors on Science and Technology and the administration's former AI czar, a day after the government ordered Anthropic to suspend worldwide access to Fable 5 and Mythos 5 under export control authorities.

In a post on X, Sacks said the administration acted "reluctantly" and only after Anthropic CEO Dario Amodei declined requests to patch the vulnerability or temporarily remove the model from service.

"The ball is in Anthropic's court," Sacks wrote, adding that the government would lift the restrictions once the jailbreak is addressed.

According to Sacks, the issue came to light after a trusted organization working with both Anthropic and the US government discovered a method of bypassing Fable 5's safety guardrails, allowing access to the unrestricted cyber capabilities of the underlying Mythos model.

He argued that Anthropic prioritized keeping its consumer model available rather than addressing a security risk, a position he said was inconsistent with the company's long-standing image as an AI safety leader that has advocated for strict regulation of advanced cyber-capable models.

Sacks also rejected suggestions that the export controls were linked to previous disagreements between Anthropic and the administration, emphasizing that the government values the company's technology and views the dispute as one that can be resolved quickly.

The controversy widened after Semafor reported, citing a person close to the White House, that Amazon had alerted the government to the jailbreak. The report said Amazon CEO Andy Jassy had discussed the issue with administration officials.

Amazon, which has invested billions of dollars in Anthropic and provides much of its cloud infrastructure, did not confirm the report. A company spokesperson said governments frequently seek its input on cybersecurity matters but declined to comment on specific conversations.

Anthropic has strongly disputed the government's characterization of the vulnerability.

The company maintains that the reported bypass is narrow and non-universal, essentially allowing users to ask the model to review a software codebase and identify vulnerabilities. It argues that comparable results can already be achieved using publicly available models, including OpenAI's GPT-5.5.

Anthropic has also said it has not been presented with evidence that the jailbreak enabled harmful cyber activity or significantly expanded the model's capabilities.

The company further argues that recalling a commercial AI model used by hundreds of millions of people over a limited jailbreak would create a precedent that could halt deployment of frontier AI systems across the industry.

Sacks disagreed with that assessment, arguing that any bypass capable of exposing what Anthropic itself has described as cyberweapon-level capabilities should be treated as a serious national security concern.

The dispute highlights the growing tension between AI innovation and government oversight. Rather than debating model performance or commercial competition, regulators and developers are now confronting a new question: what level of jailbreak risk is acceptable before authorities intervene to remove a frontier AI model from public use?

The outcome could establish an important precedent for how governments worldwide respond to safety vulnerabilities in increasingly powerful AI systems.