Anthropic Cuts Distillation Data Extraction by 90 to 99 Percent
Frontier labs are now treating their own model outputs as an attack surface, which puts the burden of defending reasoning traces on the labs rather than on the accounts sending the requests.
Reporting from 1 source: GIGAZINE.
Anthropic published defenses against distillation, the practice of training one model on another's outputs. The company says the reasoning process is the most valuable target, so it now summarizes that record before showing it to users. Anthropic also points to the free Chinese model GLM-5.3 and to requests tied to its Kimi series as reasons for the countermeasures.
Anthropic's anti-distillation work centers on the reasoning process, the text a model produces before its answer. The company's first step was to summarize that record before showing it, which it says cut benchmark scores from 80.4% to 71.8% when the summarized reasoning was distilled as-is. Distillers adapted with direct prompts and campaigns of more than 10,000 requests, so Anthropic says simpler safeguards could not keep up with operators running hundreds of thousands of accounts. It now relies on three methods, the first being a mechanism called retained thinking.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.