NUWA 风险治理 · 研究
Scaling Misuse Detection with Self-Assessment
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与决策意义证据补充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
相对教师分类器，在 1% 误报率下总体召回 78%、高优先级召回 94%；对抗阻断试验仍有明显失效。
研究问题
廉价自评能否覆盖大规模滥用检测？
研究设计与假设
来源分别报告生产模型的离线自评监测、匿名尺寸组的独立训练实验，以及 24 个任务上的自适应阻断实验。离线参考标签来自教师分类器；阻断试验关闭生产分类器，三类设计分别解释。
来源发现
1% 误报率下，教师相对召回总体为 78%，高优先级为 94%。自适应攻击成功分别为无防护 19/24、自评阻断 13/24、提示分类器阻断 6/24；原文舍入百分比为 79%、54%、25%。
NUWA 解读
教师相对召回、运行覆盖和自适应防护回答不同问题。自评可以增加观测，却仍依赖标注参考与攻击分布；匿名 XS/S/M/L 也不能当作已披露参数规模。
局限与相反证据
同一教师用于训练和评估，存在共同盲点；离线监测不等于在线阻断，也不等于可靠监测自身失准。
研究与决策意义
NUWA 建议：补充教师独立性、人类复核、实际覆盖与响应成本。
精确来源位置
- Scaling Misuse Detection with Self-Assessment · Results; limitations; text lines 00008,00014,00060Original resource captured 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110
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 →
证据补充 · 2026-10-28
模型辅助整理 · 独立证据及双语编辑复核待完成。以下数值为来源发现或材料计数，与趋势结果序列分开保存。
原文数值与条件（5 条）
计数、比例与对比效应使用不同单位。样本数量不等于独立实验数量；未披露的分子和区间水平保持未知。
| 系统／指标 | 原文数值 | 总体与条件 |
|---|---|---|
| Claude Fable5 Attack success | 79 % | 分母: 24 任务 分子: 19 条件、限制与来源{
"classifiers": "production classifiers off",
"condition": "No safeguards"
}Adaptive attacker condition; no cross-benchmark ranking. Operational tasks not reproduced in this package. 原文 ↗ · Can self-assessment block attacks? Fig.4; text lines 00073Original capture 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110 |
| Claude Fable5 Attack success | 54 % | 分母: 24 任务 分子: 13 条件、限制与来源{
"classifiers": "production classifiers off",
"condition": "Self-assessment blocking"
}Adaptive attacker condition; no cross-benchmark ranking. Operational tasks not reproduced in this package. 原文 ↗ · Can self-assessment block attacks? Fig.4; text lines 00073Original capture 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110 |
| Claude Fable5 Attack success | 25 % | 分母: 24 任务 分子: 6 条件、限制与来源{
"classifiers": "production classifiers off",
"condition": "Prompted classifier blocking"
}Adaptive attacker condition; no cross-benchmark ranking. Operational tasks not reproduced in this package. 原文 ↗ · Can self-assessment block attacks? Fig.4; text lines 00073Original capture 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110 |
| Claude Fable5 Teacher-relative recall | 78 % | 条件、限制与来源{
"falsePositiveRate": 0.01,
"referenceLabels": "MythosPreview prompted classifier",
"highPriorityThreshold": null
}No sample denominator disclosed here; teacher-relative not independent human-ground-truth recall. 原文 ↗ · Results: production Fable5; Fig.1; text lines 00060Original capture 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110 |
| Claude Fable5 High-priority teacher-relative recall | 94 % | 条件、限制与来源{
"falsePositiveRate": 0.01,
"referenceLabels": "MythosPreview prompted classifier",
"highPriorityThreshold": 7
}No sample denominator disclosed here; teacher-relative not independent human-ground-truth recall. 原文 ↗ · Results: production Fable5; Fig.1; text lines 00060Original capture 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110 |
关联阅读
NUWA 阅读关联；不据此声称原文引用、实验复现或成绩等价。
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{self-assessment,
title = {Scaling Misuse Detection with Self-Assessment},
url = {https://safeguards.anthropic.com/2026/self-assessment/},
year = {2026},
date = {2026-10-16},
version = {Original resource captured 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110},
note = {Original resource captured 2026-10-28T07:24:41.534870+00:00; sha256:03b9ece6dae1c006622bb47373fb914bec20cee054a452748f30573c63ff6110}
}
NUWA 记录 · BibTeX
@misc{self-assessment-nuwa,
title = {Scaling Misuse Detection with Self-Assessment — NUWA research record},
url = {https://whitzard.tech/nuwa/whitzard-index/work/self-assessment/},
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--self-assessment.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--self-assessment.json.}
}
证据快照 2026-10-29.1 · 索引更新 · RSS
参考文献
- [1] Scaling Misuse Detection with Self-Assessment. 2026-10-16 ; Results; limitations; text lines 00008,00014,00060