Unverified50% confidenceSolutionExact time
sleeper agent（潜伏代理）攻击范式指被污染的模型在常规测试和安全评估中表现正常，只有在特定触发条件下才展现被植入的行为，使标准评估检测极为困难
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
以色列斥资4650万美元干预AI舆论：生成式AI成为信息战新战场
Related Claims
UnverifiedVICE actively simulates attack surface probing of an application's externally exposed interactive interfaces, with a methodology derived from penetration testing and red team exercises rather than traditional passive compliance checklists.68% similarUnverified纯Agent系统测试面临非确定性输出挑战，只能依赖LLM评判器或人工评估67% similarUnverified基于规则过滤的Agent在对抗性测试中往往更容易被绕过，而经过对齐训练的模型具备更强的语义级意图识别能力64% similarUnverifiedPublicly available safety evaluation sets carry the risk of being included in training data and causing evaluation contamination, which includes two risks: model vendors gaming scores and malicious actors studying test sets to develop bypass variants.64% similarUnverifiedAgent tests should programmatically inspect the final state to verify the true end state of the system rather than relying on the model's self-report.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/700785API
curl https://kongchang.com/api/v1/knowledge/claims/700785MCP
get_claim(id=700785)