AI news · Anthropic · security
Anthropic says new defences cut copying of Claude by up to 99%
Rivals that train their own models on Claude’s answers now get 90 to 99 per cent less of its reasoning, Anthropic says.
Anthropic says new safeguards have cut the data that rivals harvest from Claude by about 90 to 99 per cent, in a post published on Monday 12 October 2026. The target is “distillation”: sending a model millions of prompts through networks of fake accounts, then training a cheaper model to imitate its answers.
The most valuable material for copiers is Claude’s reasoning, the thinking it writes out before it answers. Anthropic first showed users a summary of that reasoning instead of the full text. Anthropic reports that a model trained on full transcripts reached 80.4 per cent on the SWE-bench Verified coding benchmark, against 71.8 per cent for one trained on the summaries.
Anthropic has also banned tens of thousands of accounts tied to distillation campaigns and names China’s Zhipu, which it says tried to distil Claude while building its GLM-5.3 model, Reuters reports. Anthropic puts its share of honest requests wrongly blocked below 0.1 per cent and says it turned away meaningful revenue from these buyers.
Some legitimate customers had to update the code that connects their apps to Claude, Anthropic says, so teams that build on Claude’s reasoning output should check what their tools now receive.
Sources
- Anthropic: Distillation defenses. Checked
- Reuters via Business Standard: New safeguards sharply reduce AI distillation attacks, says Anthropic. Checked
Alexandre S.
Alexandre S. is a creative ad strategist and copywriter with ten years' experience in digital marketing across multiple channels.