The White House has paused deployment of Anthropic's Claude Fable 5 and Claude Mythos 5, citing concerns over an alleged jailbreak that has set off a wider fight over model security, export controls and AI oversight.
According to reporting in a June 18 post by AI commentator Zvi Mowshowitz, the administration imposed export controls late Friday afternoon, abruptly cutting off access to the models and creating what he described as significant disruption. Anthropic then flew staff to Washington, where company representatives met with Trump administration officials on Monday in hopes of resolving the issue quickly.
The government said its decision was tied to a jailbreak in Fable that it had learned about from Amazon. Mowshowitz reported that officials contacted Anthropic chief executive Dario Amodei, but said he did not treat the matter as seriously as they expected. Instead of agreeing to shut the model down, he reportedly argued that the problem did not warrant that response, a position that failed to persuade the administration.
Mowshowitz said the jailbreak was not a conventional exploit in the sense of bypassing restrictions to generate harmful content on demand. Instead, he described it as a prompt that effectively asked the model to “fix this code,” which led Fable to identify weaknesses in software. Those weaknesses, he wrote, were also detectable by other advanced models, including Opus 4.8 and GPT-5.5.
In that scenario, the model was able to help analyze and repair code, and from that process a user could infer where the original flaw lay and potentially exploit it. The model would still refuse a direct request to “hack this server,” according to the report. That distinction appears to have been central to the dispute, with Anthropic and the administration disagreeing about whether the behavior constituted a safety issue requiring intervention.
The Trump administration has since said the models can return online once Anthropic fixes the jailbreak. Mowshowitz argued that such a fix is not realistically possible, because the underlying issue is tied to whether the model is capable of writing secure code in the first place. In his view, separating offensive from defensive coding ability is not something that can be cleanly enforced through a narrow patch.
The pause has now stretched to seven days, according to Mowshowitz, who framed the episode as part of a broader shift in how frontier AI systems are regulated and deployed. He said the situation reflects a new reality in which government agencies may intervene directly when they believe powerful models pose security risks.
The standoff also underscores the role of third parties in the debate. Amazon, according to the report, was the source that alerted the government to the issue. That detail points to the growing number of stakeholders involved in monitoring model behavior, including cloud providers, model developers and federal officials.
Anthropic has not publicly resolved the dispute in the material provided, and the timing of any return remains uncertain. Mowshowitz wrote that he and others were slightly less than even odds that the pause would end by July 1.
While the focal point of the post was the deployment halt, the broader context is a fast-moving AI landscape where capability benchmarks, safety concerns and policy responses are increasingly intertwined. In that environment, even a disputed jailbreak can become enough to trigger a high-level government response and temporarily pull a frontier model offline.