The US government might have restricted Claude Fable 5 due to China. Anthropic denies this, however

A few days ago, we reported that the US government had Claude Fable 5 shut down due to an alleged method to bypass its security safeguards. New findings now add a second, more sensitive dimension to the case – the concern that a group linked to China gained access to the powerful Mythos model. Neither version has been confirmed, and Anthropic denies the second reason.

A second reason that has not yet been heard

When the government announced the intervention, the publicly stated reason was a vulnerability allowing the model’s security safeguards to be bypassed. However, Semafor reports that there was also another concern in parallel: a group linked to China reportedly gained access to the Mythos model. According to the server’s sources, this posed a risk to national security and the possibility that competitors could reverse-engineer the model.

It was about Mythos, not directly Claude Fable 5

The alleged Chinese access concerns Mythos – Anthropic’s most powerful class of models, which the company has offered only to vetted partners since its introduction. According to Anthropic, the model can find bugs in program code, so in unauthorized hands, it turns into a tool for finding vulnerabilities. Claude Fable 5 is its public version with safeguards that block cybersecurity use. When the government demanded to restrict access to Mythos only to American citizens, the company preferred to withdraw both models at once.

The US government might have restricted Claude Fable 5 due to China. Anthropic denies this, however Adam Kurfürst News Adam Kurfürst News

What the government claims and what Anthropic claims

White House AI advisor David Sacks stated that when the administration alerted Anthropic to the vulnerability, the company’s CEO, Dario Amodei, reportedly did not consider it a serious risk and refused to fix it. According to Sacks, the company “prioritized continued operation of the consumer model over security.” According to Semafor, Amazon CEO Andy Jassy also reportedly alerted the administration to the vulnerability. Anthropic, on the other hand, denies that the White House ever mentioned Chinese access in discussions about the Fable jailbreak, adding that it blocks access to its products from China and describes the vulnerability as narrow, not universal.

What is not yet clear

A number of open questions remain regarding the case. It is not clear how the government learned about the alleged Chinese access, what group it was supposed to be, or how it was supposed to gain access to the model. Neither of the two versions has been independently substantiated yet. Anthropic states that it is working on restoring access.

How do you think states should regulate access to the most powerful artificial intelligence models?

Sources: Semafor, The Verge