Anthropic's Fable/Mythos 5 ranks number one in the world for model intelligence, widening the US-China gap. The gap may widen further because of "anti-distillation" features, and the models are now under US export control, which has shuttered access to the advanced models.
Late Friday, the US government banned foreign governments, companies, and individuals from using Anthropic's Fable 5 and Mythos 5 models after researchers at Amazon demonstrated to the Trump administration that some safeguards on Fable could be circumvented.
People familiar with what's happening inside the Trump administration told The Wall Street Journal that Anthropic sent top officials to the White House and held calls to resolve software vulnerabilities, including the alleged ability to 'jailbreak' the model.
Anthropic's top security staff, including Nicholas Carlini, Logan Graham, and Dave Orr, were sent to Washington on Saturday to speak with senior US officials, including Commerce Secretary Howard Lutnick, National Cyber Director Sean Cairncross, and Treasury Secretary Scott Bessent. The move by the frontier AI lab aims to resolve vulnerabilities exposed by Amazon researchers.
More color from WSJ:
People close to the company and the administration said both parties are interested in resolving the issue and restoring access to the cutting-edge models, but it isn't clear what a solution would entail.
Anthropic technical experts and government security researchers coming together was seen by some administration officials as a key step toward a compromise.
The weekend discussions continue months of tension between the administration and one of America's leading AI labs over how new, cutting-edge technologies are used and regulated. The Trump administration has recently taken more steps to control the fast-evolving industry.
A Sunday letter by cybersecurity experts urged the Trump administration to lift the restrictions on the models, warning that such a move could hurt U.S. cyber defenses, create market uncertainty, and weaken America's AI leadership.
However, Jefferies analysts said quite the opposite, noting that "anti-distillation" features and US export control, "which could make it harder for open-source (Chinese) models to catch up."
"US models are improving at a faster pace likely due to compute advantage, but anti- distillation and US export control are new negatives for China AI," the analysts said.
More from Jefferies:
Open-source models (mostly Chinese) may find it harder to improve given new anti- distillation features and US export control. More importantly, Anthropic introduced anti- distillation features on Fable 5. If Fable 5 detects suspicious distillation activities, it would downgrade the model to Opus 4.8 and notify users. While this seems to be targeting Chinese AI development, we believe this would set back open source progress if all closed-source model developers follow suit. Moreover, the US has imposed emergency export control on Fable 5, barring foreigners from using them (including foreign employees of US companies), given loopholes in the cybersecurity safeguards. However, since Anthropic has no tools to limit the use to US nationals only (ie, ID checks?), it has suspended both Fable 5 and Mythos 5 globally until it could come up with a way to enforce that export control.
Did the Trump administration overreact to Amazon's findings about Fable 5? Or is it really about slowing China's open source model progress?