441e85161ClaimAnthropic says new safeguards cut the data stolen from Claude in distillation attacks by up to 99%AgreeDisagree
441e85161ClaimAnthropic says its October 2026 anti-distillation defenses cut the data distillers harvest from Claude by 90 to 99%AgreeDisagree
441e85161ClaimAnthropic said in October 2026 that its anti-distillation defenses cut the reasoning tokens distillers extract from Claude by 10 to 100 timesAgreeDisagree
441e85161ClaimAnthropic said in October 2026 that Zhipu switched from distilling Claude Fable 5 to an older Claude model and another lab's model after Claude Fable 5's cyber safeguards degraded its attacksAgreeDisagree
441e85161ClaimThe US Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) assessed Zhipu's GLM-5.3 as the most cyber-capable open-weights AI model released as of October 2026AgreeDisagree
441e85161ClaimAnthropic alleged in October 2026 that Chinese AI lab Zhipu distilled cyber capabilities from leading US AI models while developing GLM-5.3AgreeDisagree
441e85161ClaimAnthropic said in October 2026 that requests to Moonshot AI's Kimi models were secretly forwarded to Claude as part of distillation effortsAgreeDisagree
441e85161ClaimAnthropic's two-stage anti-distillation classifier blocks attempts to extract Claude's reasoning with a false positive rate under 0.1%AgreeDisagree
441e85161ClaimAnthropic reported in February 2026 that one Claude distillation network managed more than 20,000 fraudulent accounts at onceAgreeDisagree
441e85161ClaimAnthropic makes Claude Sonnet 5.5 reasoning transcripts accessible only to the organization whose request produced themAgreeDisagree
441e85161ClaimAs of October 2026, Anthropic had banned tens of thousands of distillation-linked Claude accounts, often within two hours of a distillation campaign startingAgreeDisagree
441e85161ClaimSince September 24, 2026, Anthropic has charged for Claude API requests blocked by its anti-distillation safeguardsAgreeDisagree
441e85161ClaimIn Anthropic testing, a pretrained Claude model trained on full Claude Opus 4.8 reasoning transcripts scored 80.4% on SWE-bench Verified, versus 71.8% when trained on summarized transcriptsAgreeDisagree
441e85161ClaimAI model distillation from competitor outputs constitutes intellectual property theftAgreeDisagree