一句话定义： AI safety and governance covers frontier risk, monitoring, evaluation, policy, and risk management frameworks.
页面状态
- 状态：
source-backed - 来源数量：16
- 更新方式：由 source-crawler 资料池生成，人工/LLM 综合写入
当前综合。
这是什么
AI Safety, Risk, and Governance 是 AI Wiki 中的一个长期知识节点。它不是一次性新闻，而是持续汇集官方博客、工程实践、newsletter、benchmark 和人物观点的主题页。
当前综合
- 这个主题已经有足够材料支撑第一版 wiki 页，适合继续把来源拆成概念、产品、人物和争议子页。
- 当前页面的结论应优先来自一手来源和研究/评测来源；专家观点可用于解释趋势，但不应替代原始事实。
为什么值得关注
这个主题同时出现在 16 条资料中，说明它已经跨越单篇文章，成为一个需要持续跟踪的知识簇。关键词包括：AI safety、frontier risk、AI governance、AI RMF、monitorability、MITRE ATLAS、OWASP LLM Top 10、LLM security、prompt injection、AI threat modeling。
近期信号
- The US is advancing AI safety through state and federal action：来自 OpenAI News，约 10141 字符。
- AI Safety vs AI Security: Key Differences Explained：来自 TrueFoundry Blog，约 13426 字符。
- Mythos Proves AI Safety Can No Longer Live Inside the Model：来自 grith.ai link target，约 16063 字符。
- Build an agentic AI safety pipeline with Runpod Flash and Granite Guard…：来自 RunPod Blog，约 12913 字符。
- LWiAI Podcast #243 - GPT 5.5, DeepSeek V4, AI safety sabotage：来自 Last Week in AI，约 5927 字符。
- Evaluating whether AI models would sabotage AI safety research：来自 www.aisi.gov.uk link target，约 9192 字符。
关键问题
- What is frontier AI risk? Frontier AI risk refers to risks from highly capable AI systems, including misuse, loss of control, and societal impacts.
- Why include governance? Governance connects technical evaluation with deployment rules, accountability, and public policy.
- Where do security standards fit? Security standards such as MITRE ATLAS and OWASP LLM Top 10 turn abstract AI risk into concrete attack patterns, controls, and review checklists.
待追踪问题
- 哪些来源是一手事实，哪些只是围绕 AI safety 的二次解读？
- 这个主题的证据是否足以支撑对比页、指南页或 newsletter 选题？
来源覆盖
当前页面引用了 16 条资料，主要来自：Anthropic News 1 篇、ElevenLabs Blog 1 篇、Last Week in AI 1 篇、OpenAI Alignment 1 篇、OpenAI News 1 篇、RunPod Blog 1 篇、TrueFoundry Blog 1 篇、UK AI Security Institute Research 1 篇。
证据类型
| 类型 | 数量 | 阅读建议 |
|---|---|---|
| 一手来源 | 4 | 官方或研究机构来源，适合支撑模型发布、方法、产品和政策相关事实。 |
| 发现信号 | 1 | 适合发现新主题和补充背景，重要事实应回到一手来源核对。 |
| 研究与评测来源 | 2 | 适合支撑比较、评测框架、安全分析和数据趋势判断。 |
| 背景资料 | 9 | 可作为补充上下文，阅读时需要留意发布时间和来源权威性。 |
来源列表
一手来源
- The US is advancing AI safety through state and federal action — OpenAI News，约 10141 字符
- Introducing the OpenAI Safety Fellowship — OpenAI Alignment，约 2170 字符
- Australian government and Anthropic sign MOU for AI safety and research — Anthropic News，约 6642 字符
- ElevenLabs and UK Government partner on voice AI safety research — ElevenLabs Blog，约 3119 字符
研究与评测来源
- Evaluating whether AI models would sabotage AI safety research — UK AI Security Institute Research，约 1987 字符
- AI Agent Standards Initiative — US AI Safety Institute，约 1366 字符
发现信号
- LWiAI Podcast #243 - GPT 5.5, DeepSeek V4, AI safety sabotage — Last Week in AI，约 5927 字符
背景资料
- AI Safety vs AI Security: Key Differences Explained — TrueFoundry Blog，约 13426 字符
- Mythos Proves AI Safety Can No Longer Live Inside the Model — grith.ai link target，约 16063 字符
- Build an agentic AI safety pipeline with Runpod Flash and Granite Guardian 4.1 — RunPod Blog，约 12913 字符
- Evaluating whether AI models would sabotage AI safety research — www.aisi.gov.uk link target，约 9192 字符
- Alice Powers the AI Safety Flywheel with NVIDIA — www.activefence.com link target，约 12271 字符
- The cultural variable: testing AI safety across 12 countries | Centific — www.centific.com link target，约 8456 字符
- Tracking and Debugging AI Safety Evaluations with Inspect AI and MLflow | MLflow — mlflow.org link target，约 6625 字符
- Global Standards, Local Ground Truths: Piloting Multilingual, Multimodal AI Safety Understanding in APAC - MLCommons — mlcommons.org link target，约 8666 字符
- The AI safety illusion: why current safety datasets fool us on model safety — labelbox.com link target，约 15522 字符
相关页面
- MITRE ATLAS
- OWASP LLM Top 10
- NIST AI Resource Center
- METR