China may have accessed Mythos
▼ Summary
– The White House imposed export restrictions on Anthropic’s Mythos partly due to fears it was accessed by a China-linked group, posing a national security risk.
– China could reverse engineer the model through distillation, training a “student” AI to replicate its behavior.
– The White House has not confirmed the report, and Trump advisor David Sacks focused on a jailbreak claim instead, which Anthropic denied.
– Anthropic stated the government did not mention China during export control discussions, and the company has not commented further.
– This would not be the first breach of Mythos, as a Discord group previously accessed it for two weeks before Anthropic cut off access.
A new report from Semafor reveals that the White House’s decision to impose export restrictions on Anthropic’s Mythos was partly fueled by concerns that a group linked to China had accessed the advanced AI model. If the Chinese government indeed obtained Mythos 5 or Fable 5, the implications for national security would be severe. Officials worry that Beijing could reverse engineer the model through distillation, a process where a smaller “student” AI learns from a more powerful one to mimic its capabilities.
The White House has not officially confirmed these claims. In a post on X, Trump advisor David Sacks sidestepped the China angle entirely, instead highlighting an alleged vulnerability in Fable and Mythos that allowed them to be jailbroken,a charge Anthropic has strongly denied. Anthropic has not responded to requests for comment, though a spokesperson told Semafor that the government never raised the issue of China during discussions about export controls.
If Mythos was indeed accessed by Chinese state actors, it would mark another embarrassing breach for Anthropic’s most secretive model. The company has long argued that Mythos is too dangerous and powerful for public release. Yet a Discord group reportedly managed to access it for two full weeks before Anthropic detected the intrusion and cut off their connection. This incident underscores the challenges of securing cutting-edge AI from determined adversaries.
(Source: The Verge)
