Unverified50% confidenceFactExact time
Safety alignment is not a one-time, irreversible property; fine-tuning, prompt injection, or adversarial inputs can all weaken or bypass this layer of constraint.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified提示注入目前没有一劳永逸的解决方案，护栏只能降低风险而非根除，在敏感操作场景中需叠加权限控制与人工审核等纵深防御措施79% similarUnverified大模型对指令的敏感度呈非线性分布——主干指令的微小改动可能引发全局行为漂移，而在末尾追加具体的边界条件则相对安全76% similarUnverifiedIn the parametric channel, when the constraint sensitivity oscillates, the stiffness modulation of the dual update and the damping modulation of the primal update cancel each other, producing no resonant excitation, and the system has no resonant forcing at any depth.75% similarUnverifiedPutting a spending cap in the system prompt is a soft constraint and cannot provide safety guarantees under adversarial inputs (such as prompt injection).73% similarUnverifiedNaïve confidence-stratified comparisons are unreliable because uncontrolled detection strength can confound conclusions.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1100266API
curl https://kongchang.com/api/v1/knowledge/claims/1100266MCP
get_claim(id=1100266)