Anthropic Details Distillation Defenses as US Threatens Sanctions on Moonshot AI

Anthropic published a detailed account of its defenses against distillation attacks on October 12, 2026, describing a new 'preserved thinking' feature in Claude Fable 5.1 and Opus 5.5 that blocks API accounts created on or after August 31, 2026 from editing Claude's prior context. The company reports its combined countermeasures, which also include blocking specific extraction techniques such as asking Claude to translate its reasoning into another language or write it in special tags, have reduced harvested data by around 90 to 99 percent. Anthropic states it has turned away meaningful revenue and trillions of tokens' worth of paying traffic from suspected distillers.

The defensive publication follows US government accusations in July 2026 that Moonshot AI's Kimi K3 model was trained using large-scale distillation of Anthropic's Claude Fable 5. White House OSTP director Michael Kratsios stated the administration has evidence tying Kimi K3 to a distillation campaign against Fable 5, and alleged Moonshot built in-house infrastructure to rotate access methods and evade detection. Treasury Secretary Scott Bessent warned that sanctions and Entity List designations against Moonshot AI are under consideration, and claimed watermarks of US-made models have been found in multiple Chinese-made models. The accusations also allege Moonshot used restricted Nvidia chips acquired via Thailand.

Multiple AI experts publicly called the US accusations political and reckless, arguing that AI outputs carry no copyright protection, making distillation legally distinct from intellectual property theft. A Moonshot employee pointed to the timeline, noting Kimi K3 launched only 15 days after Fable 5 went public on July 1, and wrote sarcastically that training a frontier model in 15 days would be 'Guinness World Record stuff'. Skeptics also noted that K3 testing reportedly predated Fable 5's release.

Dean Ball, writing after joining OpenAI, said distillation from US frontier models has long been part of the explanation for Chinese models performing close to US models on benchmarks, and that his time at OpenAI led him to revise that estimate upward. Ball also noted that DeepSeek's reasoning style now closely resembles Google's Astra, describing the similarity in terse reasoning style as unlikely to be coincidental.

A US national science and technology memorandum from April 2026 had already characterized foreign entities, particularly Chinese, as distilling US frontier AI at industrial scale. A congressional committee cited roughly 24,000 accounts and over 16 million interactions in an investigation of development tools incorporating Chinese-made models.