Recently, the company Anthropic released what it calls the strongest Mythos-level large model to date, Claude Fable5. With its overwhelming performance in programming and complex task handling, the model quickly became the focus of the tech community. However, just 72 hours later, this highly anticipated AI product was subjected to a "global ban" by the US government, forcing it to shut down all services.
This ban is unprecedented in its intensity. The US government, citing national security, not only prohibited individuals and institutions outside the United States from accessing the model but also explicitly stated that foreign individuals within the US (including foreign employees of Anthropic) were also banned from using it. Due to the technical difficulty of accurately and quickly identifying the nationality of each user, Anthropic ultimately opted for a "one-size-fits-all" approach, suspending all services of Claude Fable5 and its related model, Claude Mythos5. Requests from affected users will automatically be redirected to the previous version.
The incident was triggered by a major vulnerability that was discovered. According to reports, this vulnerability could bypass the model's built-in safety measures, enabling what is known as "jailbreaking." Although Anthropic stated that this was a narrow and non-general jailbreak method, and had been tested for thousands of hours through red team exercises before release, the US government took an unusually firm stance. Officials believe that since Claude Fable5 has processing capabilities close to "network weapon level," once the safety measures are breached, the consequences would be unpredictable. Therefore, they demanded the company immediately fix the vulnerability or take the product off the market.
Both sides have been locked in a fierce standoff over the severity of the vulnerability. Anthropic argued that halting a product used by hundreds of millions of users due to a simple, non-general vulnerability would set a very negative precedent for the industry. However, insiders revealed that the government's concerns went far beyond that. There are reports that the government suspects organizations related to China have already exploited this vulnerability to obtain the full capabilities of the model, posing risks of reverse engineering or model distillation.
