Unverified50% confidenceCautionExact time

水印方案在开源模型上彻底失效的根本原因之一是哈希密钥无法保密，一旦绿/红名单划分规则被逆向推断，攻击者可主动规避绿名单词汇

1

Sources

50%

Confidence

Long-term

Relevance

7/16/2026

First Seen

Sources

AI Text Detection and Privacy Protection: The Dual Dilemma of LLM Security

Related Claims

Unverified水印方案的有效性依赖模型提供方对哈希密钥的保密，这导致水印在开源模型上彻底失效79% similar VerifiedOnce open-source model weights are made public, attackers can remove a model's safety restrictions through fine-tuning.60% similar UnverifiedMatt Green推断OpenAI与Anthropic极可能都使用单一全局密钥加密所有客户端推理数据，属于密钥管理反模式58% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/533409

API

curl https://kongchang.com/api/v1/knowledge/claims/533409

MCP

get_claim(id=533409)