U.S. restrictions on foreign access to Anthropic’s latest artificial intelligence models were prompted in part by concerns that a China-linked entity may have gained access to the technology, according to a report by U.S. online media outlet Semafor citing sources familiar with the matter.
According to the report published June 13, the Trump administration directed Anthropic on June 12 to restrict access to its cybersecurity-focused AI model, Mythos 5, and its consumer version, Fable 5, to U.S. citizens only. The order also prevents foreign employees at Anthropic from using the models.
Anthropic had previously limited access to Mythos, which was released in April, to a select group of companies that could use the model to identify and address security vulnerabilities before they could be exploited by hackers. The model’s advanced ability to detect bugs in computer code has raised concerns that it could also be used for malicious hacking activities.
It remains unclear which China-related entity may have accessed Mythos or how such access may have occurred. Anthropic currently prohibits users in China from accessing its products. A company spokesperson said the White House did not raise concerns about Chinese access during discussions regarding export restrictions and potential jailbreaking of the model.
Semafor reported that access to Mythos by the Chinese government could pose a national security risk to the United States, as China could attempt to reverse-engineer and replicate the model through a process known as distillation.
Anthropic said in a public notice that it was required to immediately suspend access to Fable 5 and Mythos 5 in response to government export-control directives citing national security concerns.
The company said the government order did not specify the exact nature of those concerns but added that officials appeared to have become aware of methods for bypassing, or “jailbreaking,” Fable 5. Jailbreaking refers to techniques used to circumvent or remove an AI model’s built-in safety controls.
Anthropic said it reviewed demonstrations of the reported jailbreaking technique and found that it was used to identify only a small number of previously known vulnerabilities.
“These vulnerabilities appear relatively simple, and we confirmed that other publicly available models can identify them without any jailbreak techniques,” the company said, suggesting that some government concerns regarding national security risks may be overstated.
David Sacks, the White House’s AI and cryptocurrency policy adviser, said in a post on X that Anthropic had been warned about the possibility of Fable 5 being jailbroken. He claimed that Anthropic CEO Dario Amodei declined to address the issue, arguing that the vulnerability did not pose a serious risk.
Sacks also denied that the administration’s action was influenced by previous disagreements with Anthropic over AI regulation and the use of artificial intelligence in autonomous weapons systems. The U.S. Department of Defense designated Anthropic as a “supply chain threat” in February.
A source close to the White House told reporters that Amazon alerted the administration to the jailbreaking issue and that Amazon CEO Andy Jassy had discussed the matter with government officials.
Meanwhile, U.S. technology publication The Information reported that the Trump administration does not currently plan to extend similar restrictions on foreign access to AI models developed by other companies, including OpenAI.