Open-Source and Paid Tools Claim to Strip Anthropic's New Claude Text Watermark — Unverified

What Happened

Days after Anthropic began watermarking text generated by Claude, multiple "watermark remover" tools have appeared online, as reported by BleepingComputer. These include at least one open-source project that has already gathered over 4,500 GitHub stars, as well as paid services marketed for AI-detection evasion.

Critically, none of these tools' claims of successfully defeating Claude's text watermark can be independently verified — Anthropic has not released a public detector for the watermark, so there is currently no way to confirm or refute whether these removal techniques actually work.

Why It Matters for Defenders

AI content watermarking is increasingly positioned as a signal for provenance and authenticity — used in academic integrity checks, content moderation, disinformation triage, and increasingly in fraud and social-engineering investigations. A credible (or even just widely-adopted) evasion tool undermines confidence in that signal regardless of whether it truly works, because defenders and detection vendors downstream may have no reliable way to know if watermark-based attribution has been tampered with.

Organizations that rely on AI-content watermarking as part of any trust or verification pipeline — including those evaluating LLM-generated phishing, disinformation, or fraudulent submissions — should treat watermark presence/absence as a weak, unverified signal rather than ground truth.

What to Watch For / Do Now

• Do not treat the presence or absence of an AI watermark as authoritative proof of AI-generated content in any investigative or moderation workflow; corroborate with other indicators.
• Track the popularity and adoption of watermark-evasion tools (e.g., GitHub star counts, forks, paid service uptake) as a rough proxy for how much erosion of trust in watermarking is occurring in the wild.
• Watch for Anthropic or other vendors to release official detectors — until then, any claims of watermark removal (or watermark detection) should be treated as unverified.
• For teams building anti-abuse or content-provenance detections, avoid hard-coding assumptions about watermark robustness into automated decisioning until independent verification is possible.

Developing Story

This is a fast-moving, unverified situation — no detector exists yet to confirm whether any of these tools actually defeat Anthropic's watermark. We will monitor for updates. Read the original reporting at BleepingComputer.