Anthropic says its defenses cut Claude distillation harvesting 90–99%
Its Oct 12 report: a near-zero false-positive classifier, mass bans, preserved thinking that breaks new accounts' context edits, and billing for blocked calls.
Copy markdown
Reasoning distillers can't siphon anymore
Anthropic's Oct 12 report says layered defenses — preserved thinking, a two-stage classifier, and bans — cut the reasoning distillers extract on its platform by 10–100×, roughly 90–99% less harvested. The target: covert, industrial-scale harvesting of Claude's chain-of-thought to train rival models.
Preserved thinking breaks context edits for new accounts
If your API account was created on or after Aug 31, 2026, you can no longer edit Claude's prior reasoning context, and weaker older models can't read it — enough of a change that some teams had to rewrite integrations. Sonnet 5.5 also walls reasoning transcripts to the org that generated them.
You now pay for blocked calls
Since Sept 24, requests Anthropic's safeguards block in set categories — distillation among them — are billable, so a flagged call still burns tokens. The classifier's false-positive rate is under 0.1%, but benign high-volume traffic can still trip it.
The scale behind the clampdown
Anthropic's threat reporting ties about 200M distillation-style exchanges to five campaigns, the largest (~151M exchanges, May–July 2026) reportedly linked to Alibaba's Qwen. It says it has banned tens of thousands of accounts, often within two hours of a campaign starting.