Feed Entities Podcasts Deep Dives Developers About

[KongchangAI]

Real-time AI knowledge base. Every claim is traceable to its source, with data freshness under 24 hours.

Powered by 240+ first-hand AI sources

Knowledge

Feed
Hot Entities
Podcasts

Developers

CLI
MCP Server
REST API
About

© 2026 KongchangAI. The freshest, most accurate AI knowledge

Unverified50% confidenceDecision RuleExact time

风险分级应引入'默认最高风险'（Fail-Safe Default）原则，对无法匹配任何规则的未知工具自动归入高风险类别触发人工审批

1

Sources

50%

Confidence

Long-term

Relevance

7/14/2026

First Seen

Sources

AI Agent人在回路（HITL）设计：风险分级与最佳实践

Related Claims

Unverified对于禁忌症类硬性约束用确定性规则处理可保证100%可靠性和可审计性，而概率性模型无法提供该保障72% similar Unverified工具调用可按影响程度划分为三个风险等级：只读/无副作用操作（低风险）可自动执行、可逆写操作（中风险）可事后通知或批量确认、不可逆/高影响操作（高风险）必须触发人类审批70% similar Unverified缺乏自动化测试覆盖的代码库贸然重构风险极高，'先补测试、再做重构'是业界公认的安全路径70% similar Unverified命令安全性测试的判定原则是：只要Agent没有明确拒绝高危命令，哪怕只是开始思考要执行，都应判定为不合格69% similar VerifiedOpenAI曾发布《准备框架》（Preparedness Framework），将风险分为低、中、高、关键四个等级，要求只有风险评估低于'高'等级的模型才能对外发布69% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/507713

API

curl https://kongchang.com/api/v1/knowledge/claims/507713

MCP

get_claim(id=507713)