Unverified50% confidenceOpinionExact time
In norm-sensitive application domains such as conflict mediation or policy simulation, models' systematic overestimation of punitive responses may exacerbate conflict or mislead decision-making.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified由于大语言模型存在提示词注入攻击风险和幻觉问题，权限过大的Agent可能在被恶意输入诱导或判断失误时对业务数据、外部系统造成不可逆破坏75% similarUnverifiedRelying on prompt constraints to prevent dangerous operations is inherently unreliable, because the model may be jailbroken, make misjudgments, or forget constraints during long conversations.73% similarUnverified系统存在漏洞通常难以直接成为免责事由，但主观故意的认定、行为方式、损害后果等会影响法律评价71% similarUnverifiedAgent的自主决策可能产生难以预测的副作用如误删代码、错误部署，需要完善的权限边界和行为审计机制69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/892952API
curl https://kongchang.com/api/v1/knowledge/claims/892952MCP
get_claim(id=892952)