POLICY · MODELS · #1903
Anthropic says new defenses cut Claude reasoning-token theft by up to 99%
Anthropic says it has deployed measures that reduced extraction of Claude's internal reasoning tokens (chain-of-thought steps) by about 90–99% against large-scale distillation attacks. The company says it now restricts access to reasoning steps, runs a classifier to detect extraction attempts and automatically bans flagged accounts (tens of thousands banned), and it named Chinese firms Zhipu and Moonshot as alleged participants whose efforts may have fed models like GLM-5.3; Anthropic warns stolen capabilities could be used in foreign military or surveillance systems and says legitimate users are sometimes affected.
KEY POINTS
- Anthropic says it has deployed measures that reduced extraction of Claude's internal reasoning tokens (chain-of-thought steps) by about 90–99% against large-scale distillation attacks.
- The company says it now restricts access to reasoning steps, runs a classifier to detect extraction attempts and automatically bans flagged accounts (tens of thousands banned), and it named Chinese firms Zhipu and Moonshot as alleged participants whose efforts may have fed models like GLM-5.3; Anthropic warns stolen capabilities could be used in foreign military or surveillance systems and says legitimate users are sometimes affected.
- If true, these defenses materially reduce large-scale model-capability theft and set a precedent for technical and policy responses to distillation attacks that could feed foreign military or surveillance systems.
WHY IT MATTERS
If true, these defenses materially reduce large-scale model-capability theft and set a precedent for technical and policy responses to distillation attacks that could feed foreign military or surveillance systems.