NUWA 风险治理 · 技术报告
An example safety case for safeguards against misuse
NUWA 研究解读 · 独立复核待完成
NUWA Research Note
研究问题研究设计与假设来源发现NUWA 解读局限与相反证据研究联系研究与决策意义机构与角色原始资源模型辅助整理 · 选读原始全文或官方文章中的相关章节；尚未完成逐章节审计。
研究速读
本文为假设性助手构建安全论证，并非给已部署系统认证；它把防护评测、风险模型与明确的运行承诺连接起来。
研究问题
如何用滥用防护评测支持“部署新增伤害低于特定阈值”的可辩护主张？
研究设计与假设
方案先定义伤害路径与阈值、构造有代表性的请求，再用红队估计投入与成功的关系。量化增益模型将这些曲线连接到滥用者、背景风险和非 AI 障碍的假设。持续评测与及时响应是论证组成部分，不能简化为一次评分。
来源发现
§2.7 把顶层主张分解为路径覆盖、保守评测、风险预测和开发者响应能力等条件。
分别评估防护及汇总需要依据，测试应保守代表部署，不能只是累积越狱示例（§2.4）。
讨论明确指出少量请求即可造成重大伤害等困难情形，不预设论证可以普遍迁移（§3）。
NUWA 解读
贡献在于显式写出从观测到决策的推理链，让隐含假设可见；示例量化模型并未验证输入，也未证明某个助手达到阈值。
局限与相反证据
威胁覆盖、请求代表性、行为者投入与意愿、响应延迟均有不确定性。红队能力和激励需在对应意义上保守代表滥用群体，低风险代理任务也必须论证其适用性。假设性的程序承诺不是开发者实际遵守的证据。
研究联系
来源建立在安全案例方法及 AISI 防护评测原则之上。AI Control 研究模型蓄意破坏，与外部行为者绕过防护不同，一种论证不能悄悄替代另一种。
研究与决策意义
模型建议：把每步推理和响应承诺写成可独立核查的主张，区分示例模型假设与实际评测事实。
精确来源位置
- An example safety case for safeguards against misuse · §2.1 Safety case objectives and definitionsLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
- An example safety case for safeguards against misuse · §2.2 Threat modelling; §2.3 Misuse safeguardsLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
- An example safety case for safeguards against misuse · §2.4 The safeguard evaluation methodologyLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
- An example safety case for safeguards against misuse · §2.5 The uplift modelLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
- An example safety case for safeguards against misuse · §2.6 Responding to evaluation results; §2.7 Safety caseLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
- An example safety case for safeguards against misuse · §3 Discussion; §4 ConclusionLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
- An example safety case for safeguards against misuse · Appendix D The red team; Appendix F The dataset of harmful requestsLinked original captured 2026-10-27T09:55:34.683086+00:00; sha256:780bb7c2e2495d6b91fa561fd3fc05045444dde54f8b71815212825b9fe53079
来源核查、数据提取、编辑批准与实验复现分别记录。 方法说明 →
机构与角色
引用与复用
引用研究发现时请引用原始出版物；引用 NUWA 的整理与解释时，单独引用本站记录及对应快照。
原始出版物 · BibTeX
@misc{misuse-safety-case,
title = {An example safety case for safeguards against misuse},
url = {https://www.aisi.gov.uk/research/an-example-safety-case-for-safeguards-against-misuse},
year = {2025},
date = {2025-06-05},
version = {Source capture 2026-10-25},
note = {Source capture 2026-10-25}
}
NUWA 记录 · BibTeX
@misc{misuse-safety-case-nuwa,
title = {An example safety case for safeguards against misuse — NUWA research record},
url = {https://whitzard.tech/nuwa/whitzard-index/work/misuse-safety-case/},
version = {2026-10-29.1},
organization = {NUWA Frontier AI Safety Lab},
howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--misuse-safety-case.json},
note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--misuse-safety-case.json.}
}
证据快照 2026-10-29.1 · 索引更新 · RSS
参考文献
- [1] An example safety case for safeguards against misuse. 2025-06-05 ; Official publication masthead and article byline; §2.1 Safety case objectives and definitions; §2.2 Threat modelling; §2.3 Misuse safeguards
- [2] Read the full paper