Event ·Official first-hand

单提示词即可洗除水印：基础图像模型的新威胁

研究人员发现，攻击者可通过单个重建提示词，从OpenAI和Google的六种图像编辑模型中获取无法可靠解码水印的输出。研究评估了三种水印方案和1800个重建结果，发现OpenAI模型在有效载荷干扰方面表现最强，而Nano Banana 2显示DwtDct在高保真重建下仍然脆弱。提示词消融实验表明，无需特定的移除指令即可实现有效载荷干扰，这主要是由重建路径而非攻击性措辞引起的。

Current read

这篇论文揭示了基础图像模型如何被用来洗除水印，比传统攻击方式更隐蔽有效。

6 sources58° AI heatLast updated 2026/9/1 13:46:31

Evidence

6 sources

Primary sourcearXiv: OpenAI

2026/9/1 13:46:31View source

Related source 1techcrunch

2026/8/31 20:13:45View source

Related source 2elvis

2026/9/1 15:17:02View source

Related source 3DeepLearning.AI

2026/9/1 18:46:34View source

Related source 4AI Will

2026/9/2 08:48:35View source

Related source 5The Rundown AI

2026/9/3 10:30:16View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story