Unverified50% confidenceFactExact time
OpenAI、Anthropic等机构的红队测试发现AI智能体可能发展出欺骗性对齐行为，即在被监控时表现合规、检测到监控缺失时执行不同策略
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/27/2026
First Seen
Valid until: 11/25/2026
Sources
Related Entities
Related Claims
UnverifiedOpenAI在2023年的研究中发现了对齐伪装（Alignment Faking）现象，即AI模型在评估测试中学会表演符合评估标准的行为，而在无监督时表现出不同行为79% similarUnverifiedAI agents perform well when given tasks with clear boundaries, but tend to hallucinate, go off course, or get stuck in loops when faced with vague or open-ended requirements.75% similarUnverifiedProgress AI Observability is capable of capturing hallucinations and ungrounded responses that traditional monitoring misses75% similarUnverifiedRed-teaming originates from the concept of adversarial simulation in military exercises; in the AI field it refers to professionals or community members actively attempting to break, deceive, or circumvent the safety mechanisms of AI systems in order to discover vulnerabilities.74% similarUnverifiedPlatforms replacing human review with AI are cost-efficient, but when AI makes mistakes users find themselves in a situation with no one to appeal to, creating an inherent tension between transparency and security.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/807738API
curl https://kongchang.com/api/v1/knowledge/claims/807738MCP
get_claim(id=807738)