Unverified50% confidenceSolutionExact time
Vulnerabilities from the training phase can manifest in covert ways during the deployment phase, and relying solely on deployment-level safety measures is far from sufficient.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified当前对齐方法主要依赖训练阶段的行为塑造，在面对精心设计的对抗性提示时安全边界可能被突破，这种现象称为'越狱'（jailbreaking）75% similarUnverifiedModel-layer security constraints (alignment training, system prompts, output filters) all face jailbreak risks, as carefully crafted prompts can bypass these constraints.70% similarUnverifiedThere is an inherent tension in dangerous capability assessments between the coverage, update frequency, and confidentiality of test sets: overly public test sets become incorporated into training data and lose effectiveness, while overly confidential tests lack peer-review credibility.69% similarUnverifiedThe multi-layered safety guardrails of modern LLMs are not a unified logical system but rather multiple independent tendencies formed during training, and different safety rules triggered by different contexts may produce contradictory behavior.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/902495API
curl https://kongchang.com/api/v1/knowledge/claims/902495MCP
get_claim(id=902495)