On October 12, 2026, Anthropic published how it is defending Claude against distillation attacks, which it describes as covert, industrial-scale campaigns to copy one model’s capabilities into another without permission.
What to know
- Anthropic says early evidence shows its methods reduced the data distillers harvest from Claude by around 90 to 99%, and that extractable reasoning tokens fell 10 to 100+ times.
- It describes three layers: controlling where reasoning can go, blocking and patching extraction attempts, and making repeated attempts harder.
- For Claude Fable 5.1 and Opus 5.5, Anthropic says “preserved thinking” blocks new API accounts created on or after August 31, 2026 from editing prior context.
- A two-stage classifier now runs on several Claude models; Anthropic says its false positive rate is below 0.1%, so a benign request can occasionally be blocked.
- Anthropic says that since September 24 it has charged for requests blocked by its safeguards in areas including distillation.
- Anthropic notes some developers must update their integrations, and expects a continuing cat-and-mouse dynamic.
Why it matters
If you build on the Claude API, check your integration against Anthropic’s post, especially if you edit earlier context in long reasoning sessions. See also our Claude tool page.
Disclosure: Anthropic’s Claude is one of the tools used to help build this site. We cover Claude models on the same basis as every other maker.
Source: Anthropic. We summarise in our own words; check the source for the full details.