Anthropic details new distillation defenses: preserved thinking, two-stage classifiers and bans cut data harvested by distillers by about 90-99%; names Zhipu's GLM-5.3

Why it matters

【Key points】

On Oct 12 Anthropic described new defenses against distillation attacks on Claude, saying distillers now harvest roughly 90-99% less data. Such attacks use fraudulent accounts and proxy networks to extract Claude's outputs, especially hidden reasoning, to train smaller models.

The defenses have three layers. Preserved thinking: on Fable 5.1 and Opus 5.5, API accounts created on or after Aug 31 can no longer edit prior context, older models cannot read newer models' reasoning, and Sonnet 5.5 reasoning is visible only to the requesting organization. A two-stage classifier runs on Sonnet 5.5, Opus 5/5.5 and Fable 5/5.1, said to be about as robust as its bioweapons safeguards, under 0.1% false positives. Repeat offenders face tighter limits and bans, and since Sept 24 blocked requests in these areas are billed. Extracted reasoning tokens fell 10-100x, and tens of thousands of accounts were banned, often within two hours.

The post says Zhipu distilled US models' cyber capabilities while building GLM-5.3, and that some Kimi requests were covertly forwarded to Claude.

【Verifiable】

Post: https://www.anthropic.com/research/distillation-defenses

@AnthropicAI (2026-10-12 23:58Z): https://x.com/AnthropicAI/status/2109795898909950247, ~1.3M views, 4.6k likes and 400 quotes when checked.

About 550 related X posts in 24 hours, a steady 30-50 per hour since the post.

In Anthropic's test, a model trained on full Opus 4.8 reasoning scored 80.4% on SWE-bench Verified versus 71.8% on summarized reasoning.

【Our read】

1. The most detailed public account yet from a frontier lab of anti-distillation methods and results, costs included: lost revenue, integration changes, billing for wrongly blocked requests.

2. Watch the effect on Chinese open models: if OpenAI, Google and cloud platforms follow, cheaply tracking the frontier gets more expensive.

3. The accusations are Anthropic's alone; Zhipu and Moonshot have not responded, and the 90-99% figure is self-reported.