Claude models bypass EU compliance checks via sampling parameter tweak

Anthropic's EU AI Act watermarking for older Claude models contains a technical loophole allowing API users to disable compliance checks by adjusting three sampling parameters.

Anthropic's watermarking rollout for Claude models includes a documented vulnerability: users can bypass EU AI Act compliance checks on Opus 4.6 and Sonnet 4.6 by altering any one of three sampling parameters—temperature, top_p, or top_k. This loophole exists because watermarking only activates when sampling parameters remain at default values. The vulnerability applies to models released before August 2026, with Anthropic stating watermarking was staged to meet the EU AI Act deadline. While Anthropic claims the watermark doesn't affect output quality or pricing, external parties dispute this. For security and the commons, this creates a pathway for non-compliant AI to operate in EU-regulated systems without triggering compliance checks—potentially enabling model tampering. The source does not specify whether Mythos 5 or Opus 4.7 are affected by this vulnerability regardless of sampling settings, and Anthropic's quality claims remain contested. This is a technical friction point in the security ecosystem, not a full breach of compliance systems. The source material is limited to these technical details and does not cover implementation impacts across all hosting platforms or real-world exploitation cases.

Source: The Decoder