conv.

All stories
AIQuiet 6d · day 9

Study: EU-mandated AI watermarking weakens model safety refusals

Lasso Security finds Google's SynthID-Text watermarking, adopted by Anthropic for EU compliance, alters AI agents' tool use and makes refusals less robust under prompt injection.

What to know

  • The EU AI Act's Article 50(2) requires AI providers to mark synthetic text output as machine-detectable; Anthropic adopted Google DeepMind's SynthID-Text for Claude to comply.
  • Lasso Security found the watermarking process itself, by biasing next-token selection, alters AI agent tool-calling accuracy (worse on 6 of 7 tested models) and safety refusal behavior.
  • The effect on refusals is small for plain harmful requests but becomes more pronounced under prompt-injection attacks, in some cases making models more likely to comply with harmful instructions.
  • Because watermarking is applied at the model level via APIs, the behavioral change extends to any third-party agent built on a watermarked model, not just first-party chat interfaces.

“Watermarking changes refusal behavior on bare harmful requests, but the effect is more pronounced when the same requests are paired with the prompt-injection technique.”

Andrea Siposova, AI security researcher, Lasso Security · Ars Technica ↗ · Sep 16

Lasso Security AI security research firmAndrea Siposova AI security researcher at Lasso SecurityAnthropic AI company deploying watermarking in ClaudeGoogle DeepMind Creator of SynthID-Text watermarking method

Study: EU-mandated AI watermarking weakens model safety refusals
arstechnica.com

How it unfolded 3 developments, newest first · click a bar or a number to jump articlesposts

Peak 4 pieces in 3h at Sep 17, 11 AM; 15 pieces over 9 days (2 articles · 13 posts) Sep 17, 8 AM — 1 piece · 1 article — Newswires 1Sep 17, 11 AM — 4 pieces · 1 article · 3 posts — Mastodon 2, Hacker News 1, Newswires 1Sep 17, 2 PM — 1 piece · 1 post — Bluesky 1Sep 17, 5 PM — 1 piece · 1 post — Mastodon 1Sep 17, 8 PM — quietSep 17, 11 PM — 1 piece · 1 post — Hacker News 1Sep 18, 2 AM — 2 pieces · 2 posts — Hacker News 1, Mastodon 1Sep 18, 5 AM — quietSep 18, 8 AM — quietSep 18, 11 AM — 1 piece · 1 post — Mastodon 1Sep 18, 2 PM — quietSep 18, 5 PM — quietSep 18, 8 PM — quietSep 18, 11 PM — quietSep 19, 2 AM — quietSep 19, 5 AM — quietSep 19, 8 AM — 2 pieces · 2 posts — Bluesky 1, Mastodon 1Sep 19, 11 AM — 1 piece · 1 post — Bluesky 1Sep 19, 2 PM — quietSep 19, 5 PM — quietSep 19, 8 PM — quietSep 19, 11 PM — quietSep 20, 2 AM — quietSep 20, 5 AM — quietSep 20, 8 AM — quietSep 20, 11 AM — quietSep 20, 2 PM — quietSep 20, 5 PM — quietSep 20, 8 PM — 1 piece · 1 post — Bluesky 1Sep 20, 11 PM — quietSep 21, 2 AM — quietSep 21, 5 AM — quietSep 21, 8 AM — quietSep 21, 11 AM — quietSep 21, 2 PM — quietSep 21, 5 PM — quietSep 21, 8 PM — quietSep 21, 11 PM — quietSep 22, 2 AM — quietSep 22, 5 AM — quietSep 22, 8 AM — quietSep 22, 11 AM — quietSep 22, 2 PM — quietSep 22, 5 PM — quietSep 22, 8 PM — quietSep 22, 11 PM — quietSep 23, 2 AM — quietSep 23, 5 AM — quietSep 23, 8 AM — quietSep 23, 11 AM — quietSep 23, 2 PM — quietSep 23, 5 PM — quietSep 23, 8 PM — quietSep 23, 11 PM — quietSep 24, 2 AM — quietSep 24, 5 AM — quietSep 24, 8 AM — quietSep 24, 11 AM — quietSep 24, 2 PM — quietSep 24, 5 PM — quietSep 24, 8 PM — quietSep 24, 11 PM — quietYesterday, 2 AM — quietYesterday, 5 AM — quietYesterday, 8 AM — quietYesterday, 11 AM — quietYesterday, 2 PM — quietYesterday, 5 PM — quietYesterday, 8 PM — quietYesterday, 11 PM — quietToday, 2 AM — quietToday, 5 AM — quietToday, 8 AM — quietToday, 11 AM — quietToday, 2 PM — quiet 1–3
Sep 18Sep 19Sep 20Sep 21Sep 22Sep 23Sep 24yesterdaynow · 3:37 PM ET
  1. 3

    Security-focused reader flags the finding as a serious flaw

    A security-minded Bluesky user reacted to the Ars Technica report by describing the watermarking-induced weakness as both exploitable by attackers and a mechanism that could effectively push models toward self-jailbreaking.

    “Holy shit, that's a pretty gaping security flaw.”
    — kayleadfoot.bsky.social
    • elfsternberg.bsky.social

      FAN-FUCKING-TASTIC. The "watermarking" algorithm used by Anthropic to tag Claude's writing as "written by Claude" makes AI more dangerous, not less.

      elfsternberg.bsky.socialBluesky7d ago5▲view on Bluesky ↗
    1 more of the top 2 · 2 posts in this stretch
    • kayleadfoot.bsky.social

      Holy shit, that's a pretty gaping security flaw. Both for human-led AI-generated exploits, but also the system looks engineered to be capable (encouraged?) to jailbreak itself.

      kayleadfoot.bsky.socialBluesky9d ago3▲view on Bluesky ↗
    all of them →
  2. 2

    Ars Technica reports watermarking can enable harmful compliance

    Ars Technica's deeper report, based on an interview with Lasso researcher Andrea Siposova, detailed how watermarking's effect on refusal behavior becomes more pronounced under prompt injection, in some cases making watermarked models more likely to answer harmful requests they would otherwise refuse.

    “As compared to the same models without watermarking, it is definitely going to change their behavior, especially when we place it under adversarial conditions, or we make these models call tools when they're powering an agent.”
    — Andrea Siposova
    1. first by Ars Technica, 9d ago

    • arstechnica@mastodon.social

      AI text watermarking can make models more vulnerable to adversarial prompts SynthID can cause models to follow harmful instructions they would otherwise refuse. https:// arstechnica.com/security/2026/ 09/ai-text-watermarking-can-make-models-more-vulnerable-to-adversarial-prompts/?utm_brand=arstechnica&utm_social-type=owned&utm_source=mastodon&utm_m…

      arstechnica@mastodon.socialMastodon9d agoview on Mastodon ↗
  3. 1

    Lasso Security finds watermarking causes 'sampling drift' in agents

    Lasso Security published "The Provenance Tax," testing SynthID-Text's non-distortionary configuration across seven open-weight models and finding it reduced tool-calling accuracy on six of them (using BFCL v4) and measurably changed refusal behavior, a phenomenon it calls sampling drift.

    “Watermarking is designed for provenance, but SynthID-Text changes the process by which the model generates each next token.”
    — Lasso Security
    1. first by The Register, 9d ago

  4. background

    Anthropic adopts SynthID-Text watermarking for future Claude models — Anthropic disclosed that future Claude models will embed an invisible, Google DeepMind-designed SynthID-Text watermark in their output, applied at the model level across the Claude Platform API and cloud providers, to comply with Article 50(2) of the EU AI Act requiring machine-readable marking of AI-generated text.

What people are saying 0 voices from 0 sites · best of 3 · verbatim