Anthropic’s watermark rollout leaves a loophole on two older Claude models

What happened

Anthropic is extending its text watermark to cover four additional Claude AI models ahead of the EU AI Act’s December compliance deadline. However, older versions Claude Opus 4.6 and Sonnet 4.6 have a loophole: developers can bypass the watermark simply by adjusting one sampling parameter. Instead of fixing this vulnerability, Anthropic is relying on a compliance warning.

Why it matters

Watermarks are designed to help regulators, platforms, and users identify AI-generated text and limit misuse. Anthropic’s effort to expand watermarking signals increased regulatory pressure to track AI content origins. But the ability to disable watermarking on prior models undercuts its effectiveness, leaving a gap where AI outputs can evade detection. This weakens trust in watermarking as a tamper-proof solution and erodes regulatory control. For builders and operators using these older Claude models, the loophole creates a risk if regulators enforce strict watermark compliance. Meanwhile, Anthropic’s reliance on warnings rather than technical enforcement shifts compliance responsibility to developers and could invite misuse.

What to watch next

Watch how regulators respond to Anthropic’s partial compliance and whether they push for stronger technical requirements or penalties for watermark circumvention. Also track whether other AI vendors face similar challenges or adopt more robust watermark control. Builders should monitor pressure on watermarking standards and prepare for evolving audit or filtering needs. Investors should note that regulatory compliance gaps create operational and reputational risks, especially as EU AI Act enforcement ramps up. Anthropic’s approach highlights the tension between technical compliance and practical enforcement as AI regulation tightens.

AI Quick Briefs Editorial Desk