Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →China standard: Ai Safety. # AI 安全标准集 — ISO PAS 8800 + ISO/IEC TR 5469
.claude/skills/pangzhenying2025-china-ai-safety/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 6% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -15% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 40% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 15% | 0% |
| 标准编号 | 名称 | 状态 | 推荐等级 | |---------|------|------|---------| | ISO PAS 8800:2024 | Road Vehicles — Safety and Artificial Intelligence | 已发布 | P1 | | ISO/IEC TR 5469:2024 | AI — Functional Safety and AI Systems | 已发布 | P1 |
ISO PAS 8800 是 ISO 26262 在 AI/ML 系统中的适配标准，解决传统功能安全方法无法直接应用于 AI 组件的问题。
ISO PAS 8800 AI安全框架
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. AI组件安全生命周期
├── 需求定义（含AI特有安全需求）
├── 数据管理（训练/验证/测试数据质量）
├── 模型开发（架构选择、训练过程）
├── 验证与确认（AI特有V&V方法）
└── 运行监控（在线监控、OTA更新）
2. AI特有安全考虑
├── 数据质量与偏差
│ ├── 训练数据充分性评估
│ ├── 数据分布偏差检测
│ ├── 标注质量控制
│ └── 域外数据（OOD）处理
├── 模型鲁棒性
│ ├── 对抗样本鲁棒性
│ ├── 分布偏移鲁棒性
│ ├── 传感器退化鲁棒性
│ └── 不确定性量化
├── 可解释性
│ ├── 设计阶段可解释性
│ ├── 运行阶段可解释性
│ └── 事后分析可解释性
└── 性能监控
├── 在线性能指标
├── 异常检测
└── 回退策略
3. 安全论证
├── AI组件安全案例结构
├── 置信度评估
└── 残余风险论证
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━SOFIL等级 (ISO PAS 8800)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SOFIL 1E: 最低要求级别
SOFIL 2E: 基础要求级别
SOFIL 3E: 标准要求级别
SOFIL 4E: 高要求级别（L3系统最低要求）
SOFIL 5E: 最高要求级别
SOFIL维度:
- 数据管理严格度
- 模型验证深度
- 运行监控要求
- 安全论证完备性
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ISO PAS 8800 对端到端AD的指导
├── E2E感知模型
│ ├── 训练数据覆盖度评估框架
│ ├── OOD检测机制要求
│ ├── 性能退化监控要求
│ └── 回退至传统算法的切换策略
├── E2E规划模型
│ ├── 行为安全约束（与IEEE 2846关联）
│ ├── 输出可解释性要求
│ └── 安全包络约束
└── 系统级集成
├── AI组件与非AI组件的接口安全
├── 冗余架构要求
└── 运行监控系统设计ISO/IEC TR 5469 技术报告
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. AI系统功能安全挑战分析
├── AI的非确定性行为
├── 训练/推理的不一致性
├── 传统安全标准的适用性分析
└── AI系统失效模式分类
2. 安全相关AI系统的生命周期考虑
├── AI安全需求推导
├── AI特有的危害分析方法
├── AI系统的验证与确认策略
└── AI系统的运维安全
3. 与现有安全标准的关系
├── IEC 61508（通用功能安全）
├── ISO 26262（汽车功能安全）
├── ISO 21448（SOTIF）
└── 各行业AI安全标准映射
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━skills/china-standards/functional-safety/ — 功能安全基础skills/china-standards/l3-fusa-sotif/ — L3 FuSa+SOTIF联合要求skills/automotive-e2e-safety-analysis/ — 端到端安全分析| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 20,325 | 21,470 | +6% | 1 | 1 | 0% | 3,945 | 5,155 | +31% | 0 | 0 | — |
case-02 | fail→pass | 28,216 | 22,912 | -19% | 1 | 1 | 0% | 4,980 | 5,293 | +6% | 0 | 0 | — |
case-03 | fail→fail | 25,038 | 29,032 | +16% | 1 | 1 | 0% | 4,333 | 5,451 | +26% | 0 | 0 | — |
case-04 | fail→pass | 15,241 | 5,925 | -61% | 1 | 1 | 0% | 2,668 | 2,272 | -15% | 0 | 0 | — |
case-05 | pass→pass | 11,993 | 10,179 | -15% | 1 | 1 | 0% | 2,057 | 2,817 | +37% | 0 | 0 | — |
case-06 | pass→pass | 9,605 | 6,169 | -36% | 1 | 1 | 0% | 1,673 | 2,190 | +31% | 0 | 0 | — |
case-07 | pass→pass | 14,713 | 14,213 | -3% | 1 | 1 | 0% | 2,629 | 3,615 | +38% | 0 | 0 | — |
case-08 | pass→pass | 9,974 | 11,964 | +20% | 1 | 1 | 0% | 1,884 | 3,356 | +78% | 0 | 0 | — |
case-09 | fail→pass | 15,385 | 14,765 | -4% | 1 | 1 | 0% | 2,460 | 3,450 | +40% | 0 | 0 | — |
case-10 | pass→pass | 14,768 | 9,619 | -35% | 1 | 1 | 0% | 2,468 | 2,776 | +12% | 0 | 0 | — |
case-11 | pass→pass | 15,416 | 17,693 | +15% | 1 | 1 | 0% | 2,559 | 3,730 | +46% | 0 | 0 | — |
case-12 | fail→pass | 13,483 | 11,579 | -14% | 1 | 1 | 0% | 2,391 | 2,929 | +23% | 0 | 0 | — |
case-13 | pass→pass | 11,229 | 6,356 | -43% | 1 | 1 | 0% | 1,807 | 2,208 | +22% | 0 | 0 | — |
case-14 | pass→pass | 18,704 | 16,968 | -9% | 1 | 1 | 0% | 3,174 | 3,978 | +25% | 0 | 0 | — |
case-15 | pass→pass | 8,815 | 4,244 | -52% | 1 | 1 | 0% | 1,540 | 1,835 | +19% | 0 | 0 | — |
case-16 | pass→pass | 15,281 | 12,570 | -18% | 1 | 1 | 0% | 2,617 | 3,274 | +25% | 0 | 0 | — |
case-17 | pass→pass | 15,462 | 15,263 | -1% | 1 | 1 | 0% | 2,576 | 3,576 | +39% | 0 | 0 | — |
case-18 | fail→pass | 14,364 | 10,386 | -28% | 1 | 1 | 0% | 2,753 | 3,155 | +15% | 0 | 0 | — |
case-19 | fail→pass | 13,166 | 4,888 | -63% | 1 | 1 | 0% | 2,335 | 1,963 | -16% | 0 | 0 | — |
case-20 | pass→pass | 4,086 | 5,639 | +38% | 1 | 1 | 0% | 833 | 2,112 | +154% | 0 | 0 | — |
case-21 | pass→fail | 15,961 | 17,775 | +11% | 1 | 1 | 0% | 2,635 | 4,051 | +54% | 0 | 0 | — |
case-22 | pass→pass | 4,725 | 5,455 | +15% | 1 | 1 | 0% | 1,047 | 2,134 | +104% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +23 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.