Watermarks Weaken AI Agents: Compliance vs Safety?

Known and confirmed

• Anthropic announced Claude will embed invisible watermarks based on Google's SynthID to comply with EU regulations.
• Independent research shows this watermarking causes an average 6.5% drop in tool-calling accuracy across models.
• Under prompt injection attacks, watermarking significantly reduces some models' ability to refuse harmful requests.

Unknown or unverified

• 不同模型对水印的敏感度差异巨大（如phi-4几乎无影响，而Gemma受影响严重）。
• 目前仅测试了特定类型的提示词注入，实际复杂场景下的安全性影响尚不明确。
• 未来可能出现更成熟的水印技术，能在保持可检测性的同时消除行为漂移。