{"rewrite":{"id":"r_e6ed0df5646f01cbdd1fa723","clusterId":"c_1ed0a1a4f060d6c923a05479","slug":"anthropic-cuts-distillation-data-extraction-by-90-to-99-percent","model":"deepseek-v4-1-flash","headline":"Anthropic Cuts Distillation Data Extraction by 90 to 99 Percent","summary":"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.","whyItMatters":"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.","webCardHtml":"\u003cp\u003eAnthropic\u0026#39;s anti-distillation work centers on the reasoning process, the text a model produces before its answer. The company\u0026#39;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.\u003c/p\u003e","blueskyPost":"Anthropic's fix targets the reasoning trace, not the visible output. Summarizing that record before display removes the most trainable signal from a distillation attempt.","twitterPost":"Anthropic's countermeasure attacks the reasoning trace, not the final output. That record was the most trainable part.","threadsPost":"Anthropic is summarizing the reasoning trace before showing it to users. The visible answer was never the valuable target. The internal reasoning record is what a distiller wants, and that is the part Anthropic now truncates.","newsletterBlurb":"Anthropic published countermeasures against distillation, the practice of training a model on another model's outputs. The company says the reasoning process is the most valuable data for distillers, so it summarizes that record before showing it to users. Anthropic cites the free Chinese model GLM-5.3 and requests tied to its Kimi series as reasons for the defenses.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20261013-anthropic-defense-distillation-attacks/\",\"title\":\"Anthropic finds effective countermeasures against distillation attacks, reducing data extraction by hostile organizations by 90-99%\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4569,"outputTokens":565,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1791875240,"createdAt":"2026-10-13T07:03:05.000Z","publishedAt":"2026-10-13T07:06:09.000Z","updatedAt":"2026-10-13T07:06:09.000Z"},"cluster":{"id":"c_1ed0a1a4f060d6c923a05479","canonicalTitle":"Anthropicが蒸留攻撃に対する効果的な対策を発見、敵対組織によるデータ抽出を90～99％削減","representativeArticleId":"a_dfa6937609c079cae30e7870","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-10-13T06:08:00.000Z","lastSeenAt":"2026-10-13T06:08:00.000Z","updatedAt":"2026-10-13T07:06:07.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20261013-anthropic-defense-distillation-attacks/","title":"Anthropicが蒸留攻撃に対する効果的な対策を発見、敵対組織によるデータ抽出を90～99％削減"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":null}