Unverified50% confidenceDecision RuleExact time
任何将原始不可信文件内容直接灌入LLM的事件响应或恶意软件分流流程都应把这些内容当作对抗性输入，在模型看到之前扫描
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
GuardBreaker攻击：用核武器提示词瘫痪AI恶意软件分析
Related Entities
Cite This Claim
Stable URI
https://kongchang.com/claim/1000248API
curl https://kongchang.com/api/v1/knowledge/claims/1000248MCP
get_claim(id=1000248)