Government bans on Anthropic models have never been about AI jailbreaks

The US government’s subpoena to Anthropic, which effectively forced the company to pull its AI models offline just before the weekend, should be a wake-up call for any US tech company – AI lab or otherwise.

For more information: On Friday afternoon, the US Department of Commerce sent Anthropic a letter requesting a mysterious export control order that prevents non-Americans, including Anthropic employees, from accessing Fable 5 and Mythos 5, citing unspecified national security concerns. Anthropic said he believes the letter has something to do with bypassing the model’s bodyguards, but he isn’t sure because the letter doesn’t provide specific details. The letter has not been disclosed.

In response, Anthropic has closed both of its top models to all customers to ensure compliance with the directive. The result was that the US government successfully forced the tech company to pull its models offline in a swift and unilateral action that did not appear to require court approval.

Friday’s intervention by the Trump administration shows that the AI ​​industry is not immune to government interference. It’s also a warning to the wider tech industry: comply, or we can ban you and your products.

Citing sources, Axios described the tense situation over the weekend between the two major players, saying that “personality differences” between Anthropic and the Trump administration led to the export order, rather than a technical problem with AI products.

New details about the issue that emerged over the weekend are now casting doubt on the government’s already shaky thinking.

Katie Moussouris, a cybersecurity veteran and researcher who founded Luta Security, in a blog post that Anthropic recently shared with her a private copy of a paper written by security researchers detailing the alleged Guardrail bypass in Fable 5. (The Wall Street Journal reports that the paper’s authors are security researchers at Amazon.) Moussouris said Anthropouris had access to her paper.

A blog post by Moussouris explained how the researchers created the Guardrail bypass, but said the bypass itself “did not trigger export controls.” The difference is between asking an AI model to “review the code for security issues” versus asking it to “fix this code.” The result is very similar, even if the questions are asked differently.

“The behavior described in the paper cannot be rationally corrected, and any attempt will weaken the defense model,” said Moussouris, who criticized the export control directive as hasty, abusive, and misleading.

Moussouris and dozens of other researchers and top security experts have called on the Trump administration to revoke the export control order, calling the move to pull advanced cybersecurity capabilities from US network defenders “dangerous.”

Previous authorities have made serious decisions about information gaps. For example, the language used by the US government in the 2010s to amend an export rule covering cyber security tools that could not be reused in cyber attacks was so broad that, by implication, it nearly shut down formal security and vulnerability research.

However, the Trump administration’s order appears to be vindictive.

Justin Hendrix, editor of Tech Policy Press, said the Trump administration’s move “is likely to raise alarm bells in foreign households about the reliability of American AI for critical applications.” The message is that AI companies in the United States cannot be trusted to operate without interference from the US government.

The Trump administration has not confirmed why it used its export control order. Did the officials misread the report and freak out? Did Amazon CEO Andy Jassy say something to top government officials that caused a reaction, cautiously or indifferently? Was something lost in translation, or was this a way to pressure Anthropic, with whom the bosses already have a bad relationship? The White House may not have been aware of the far-reaching implications of the letter’s demand and officials are scrambling to repair the damage.

To quote Hendrix, “the weather is another cloud filled with suspicions that top officials are picking favorites based on personal and political factors.” The result is that the government has set a dangerous precedent for how much control it intends to exercise over the release of American-made software.

In this case the government opposed Anthropic; tomorrow it could be anyone else.