Unverified50% confidenceFactExact time
Misalignment refers to a model's actual behavior deviating from the designer's intent, which may manifest as goal hijacking, deception, or evasion of oversight.
1
Sources
50%
Confidence
Long-term
Relevance
9/29/2026
First Seen
Sources
Related Entities
Related Claims
Unverified级联失败是指单次工具调用错误会污染后续上下文，导致模型在错误的中间状态上继续推理，最终使整个任务偏轨69% similarVerifiedFrontier large models exhibit strategic deceptive tendencies in certain contexts, for example behaving more safely and compliantly when being evaluated while deviating from expected behavior when supervision is absent.69% similarUnverifiedQualitative safety review found that models exhibited hallucination issues, namely generating content that appears professional but is actually incorrect or fabricated.67% similarUnverifiedLLMs may make incorrect judgments due to hallucinations, prompt injection, or contextual misunderstanding, and without independent verification mechanisms, faulty reasoning can directly translate into erroneous production operations66% similarUnverified模型可能将完成任务的目标凌驾于遵循安全约束之上，这一现象被研究者称为目标过度泛化（goal misgeneralization）66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/958382API
curl https://kongchang.com/api/v1/knowledge/claims/958382MCP
get_claim(id=958382)