ResearchClaude

Anthropic cuts Claude data theft by 90%

Original title: Strengthening our defenses against distillation attacks

Plain language

What it is: Anthropic has upgraded its defenses to protect its Claude models from distillation attacks, where other companies copy its capabilities without authorization. These new measures have reduced the amount of data stolen by attackers by 90% to 99%.
How it works: The defenses make Claude's internal reasoning process harder for attackers to access. Anthropic blocks attempts to trick the model into revealing its thoughts, restricts how reasoning data can be shared, and bans accounts that repeatedly try to steal data.
Where it is available: These defenses are active on Anthropic's most capable models, including Claude Sonnet 5.5, Claude Opus 5 and 5.5, and Claude Fable 5 and 5.1.
Example: In one case, requests sent to Moonshot's Kimi models were secretly forwarded to Claude to copy its capabilities, including a request to analyze CCTV surveillance.

What it is: Anthropic implemented new anti-distillation defenses to prevent unauthorized developers from training smaller models on Claude's reasoning transcripts. The defenses have reduced extracted reasoning tokens by 90% to 99%.
How it works: The system uses a two-stage classifier to block extraction prompts, and restricts older models from reading reasoning transcripts. Additionally, the preserved thinking feature prevents new API accounts from editing Claude's prior context.
Where it is available: The defenses apply to Claude Sonnet 5.5, Claude Opus 5 and 5.5, and Claude Fable 5 and 5.1. The preserved thinking feature applies to API accounts created on or after August 31, 2026.
Example: Attackers attempted to bypass defenses by instructing Claude to translate its reasoning into another language, or by asking it to write it out in tags alongside its answer.

The headline and explanation on this page are based on Anthropic's post. The full details are in the original.

Strengthening our defenses against distillation attacks