Anthropic cuts AI model distillation data harvesting by up to 99 percent
Anthropic announced that new safeguards have reduced data harvested by rival artificial intelligence developers through distillation attacks by 90 to 99 percent, prompting the company to turn away trillions of tokens in paying traffic.
Anthropic has deployed new defenses that slashed the volume of model data harvested by rival artificial intelligence developers by 90 to 99 percent, reducing extraction by a factor of 10 to 100. In implementing the protections across its Claude models, the company stated it turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.[1][2][4][5]
The company's counter-distillation measures include features such as preserved thinking in Fable 5.1 and Opus 5.5, along with tighter access controls and bans on thousands of accounts tied to extraction campaigns. Anthropic reported that a single adversarial network once operated more than 20,000 fake accounts simultaneously. Company data showed average daily reasoning tokens exposed to distillation dropped from 8,800 times a quiet-day baseline prior to activating defenses on September 19 to 135 times after.[3][4][5]
Anthropic specifically named Chinese AI developers Zhipu and Moonshot among the entities distilling Claude models. According to Anthropic, Zhipu sought to distill cyber capabilities from Claude Fable 5 while building GLM-5.3, but shifted to an older Claude model and another lab's system after safeguards degraded its attempts. While the US Center for Advancing Innovation and Standards for Super Intelligence assessed GLM-5.3 as the most cyber-capable open-weights model to date, Anthropic stated its researchers found the model's safeguards can be bypassed easily to generate security exploits.[4]
Key facts
- Anthropic reported that new safeguards reduced data harvested by rival AI distillers by 90 to 99 percent.
- Anthropic stated it turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.
- Defensive measures included preserved thinking in Fable 5.1 and Opus 5.5, tighter access controls, and bans on thousands of extraction accounts.
- Anthropic said one distillation network ran more than 20,000 fake accounts at the same time.
- Reasoning tokens exposed to successful distillation fell from an average of 8,800 times baseline before September 19 safeguards to an average of 135 times baseline afterward.
- Anthropic named Chinese labs Zhipu and Moonshot as distillers, stating Zhipu attempted to distill cyber capabilities from Claude Fable 5 for GLM-5.3 before switching to older models.
Sources · 5 sources
- GU
GURGAVIN@gurgavinPost on X ·
*ANTHROPIC: REDUCED AMOUNT OF DATA HARVESTED BY DISTILLERS BY AROUND 90 TO 99 PERCENT *ANTHROPIC: WE'VE TURNED AWAY MEANINGFUL REVENUE AND TRILLIONS OF TOKENS' WORTH OF PAYING TRAFFIC FROM DISTILLERS
Open source - CB
Crypto BriefingArticle ·
Anthropic says new defenses cut distillation data harvested from Claude by 90 to 99 percent Enhanced defenses against distillation attacks bolster AI security, ensuring model integrity and safeguarding proprietary technology advancements. The post Anthropic says new defenses cut distillation data harvested from Claude by 90 to 99 percent appeared first on Crypto Briefing .
Open source - AI
Analytics Insight@analyticsinmePost on X ·
𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜 𝐂𝐥𝐚𝐮𝐝𝐞 𝐒𝐞𝐜𝐮𝐫𝐢𝐭𝐲: 𝐃𝐚𝐭𝐚 𝐇𝐚𝐫𝐯𝐞𝐬𝐭𝐢𝐧𝐠 𝐃𝐫𝐨𝐩𝐬 𝐛𝐲 𝐔𝐩 𝐭𝐨 𝟗𝟗% Anthropic says new safeguards have reduced AI distillation attacks by up to 99%. The company reports stronger Claude AI protections, tighter access controls and bans on thousands of accounts linked to suspected extraction campaigns. The move highlights growing concerns over AI model security and competition between US and Chinese developers. #Anthropic #ClaudeAI #ArtificialIntelligence #AISecurity #AIModels #TechNews #analyticsinsight #analyticsinsightmagazine Read More 👇 https://t.co/xH2geny92g
Open source - CH
Chubby♨️@kimmonismusPost on X ·
Anthropic says new safeguards have cut the distillation that rival labs can pull out of Claude by 10 to 100 times ("around 90 to 99 percent"). Anthropic names Chinese labs Zhipu and Moonshot, and says one network once ran more than 20,000 fake accounts at a time. It remains to be seen whether GLM and Kimi will continue to release outstanding models when they can barely distill anymore.
Open source - TE
Techmeme@TechmemePost on X ·
Anthropic says new measures like preserved thinking in Fable 5.1 and Opus 5.5 have reduced data harvesting through distillation attacks by about 90% to 99% (Anthropic) (Visit Techmeme dot com for the link and full context!)
Open source