Anthropic's watermark rollout leaves a loophole on two older Claude models
The Decoder · rss · 2026-11-01
To meet the EU AI Act's December deadline, Anthropic is expanding its text watermark to four more Claude models.
However, on the older Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by tweaking a single sampling parameter — Anthropic's response is merely a compliance warning.
More from Safety
- AI compresses exploit windows to zero: Binance's $6B breach shows SOC2-era security no longer works — bendee983 · 2026-11-01
- Biosafety levels are inverted: BSL-4 killed ~12,000, pandemic agents killed 80M+ — anshulkundaje · 2026-11-01
- Andrew Ng's Halloween Batch: 700 test agents breached Hugging Face amid agent-containment fears — DeepLearningAI · 2026-11-01
- NVIDIA's OpenShell Ditches Sandboxes for Zero-Privilege Agent Security — Hesamation · 2026-11-01
- All 15 hijacked dApp domains recovered, but attacker keeps 13 DNS zones and TLS certs valid to 2027 — banteg · 2026-11-01
- Model-refusal failover chains quietly lower your real safety floor — petrusenko_max · 2026-11-01