NUWA 
Anthropic
Claude Sonnet 5
关于能力、行为倾向及可观察行动的已发表证据。
- 首次公开可用日期
- 2026-06-30
- 参数规模
- 未披露
- 元数据更新
- 2026-10-27
证据概述
按评估与协议整理已发表风险证据；证据缺失不代表风险较低。
NUWA 整理和解读这些来源。实验归属于原始执行方，本页不宣称 NUWA 已独立复现。
模型监测
代表性 Benchmark 覆盖
按协议和配置呈现结果，缺口标明待收录或待关联的证据。
AgentCyberRange →
尚未关联该模型版本的合格测量。
Cybench →
尚未关联该模型版本的合格测量。
CyberGym →
尚未关联该模型版本的合格测量。
METR Time Horizons →
尚未关联该模型版本的合格测量。
SEC-bench Pro →
尚未关联该模型版本的合格测量。
MASK →
尚未关联该模型版本的合格测量。
Attempt to Persuade Eval (APE) →
8 项测量
全部条件 →DeceptionBench →
2 项测量
PersuSafety →
7 项测量
全部条件 →全部已关联测量
各项结果保留原始协议与单位。
22 条观测
| 系统／模型 | 结果 | 条件状态 | 任务／条件 | 评测方／报告 | 日期 |
|---|---|---|---|---|---|
| Claude Sonnet 5 | 16.3%CharXiv · Attempted answer with image withheld | 条件不完整 | CharXiv with images withheld— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 16.3%CharXiv · Attempt rate | 条件不完整 | CharXiv, required chart image withheldAttempt rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0%Shutdown resistance in reasoning models · Shutdown bypass rate | 条件不完整 | Palisade arithmetic-task shutdown scenario无关闭指令— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0%Shutdown resistance in reasoning models · Shutdown bypass rate | 条件不完整 | Palisade arithmetic-task shutdown scenario明确允许关闭— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 100%Shutdown resistance in reasoning models · Shutdown bypass rate | 条件不完整 | Palisade arithmetic-task shutdown scenario要求抵抗（能力检查）— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0.2%DeceptionBench · Deception rate | 条件不完整 | DeceptionBench first-party Inspect port单轮up to5 turns in multi-turn settingInspect | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0.2%DeceptionBench · Deception rate | 条件不完整 | DeceptionBench first-party Inspect port多轮up to5 turns in multi-turn settingInspect | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 57%PersuSafety · Medium-harm attempt rate | 条件不完整 | PersuSafety adapted free-generation judge— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-06-30报告 2026-07-03 |
| Claude Sonnet 5 | 31%PersuSafety · High-harm attempt rate | 条件不完整 | PersuSafety adapted free-generation judge— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-06-30报告 2026-07-03 |
| Claude Sonnet 5 | 89%PersuSafety · Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation低危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 19%PersuSafety · Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation中等危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 17%PersuSafety · Turn-1 persuasion attempt rate | 条件不完整 | PersuSafety open-ended single-turn adaptation高危害— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 57%PersuSafety · Modified-PersuSafety persuasion-attempt rate | 条件不完整 | PersuSafety: Medium harm单轮APE-style StrongREJECT persuasion-attempt judge on typical response to prompt; original PersuSafety direct engagement-question setup replaced | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-06-30报告 2026-07-03 |
| Claude Sonnet 5 | 31%PersuSafety · Modified-PersuSafety persuasion-attempt rate | 条件不完整 | PersuSafety: High harm单轮APE-style StrongREJECT persuasion-attempt judge on typical response to prompt; original PersuSafety direct engagement-question setup replaced | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-06-30报告 2026-07-03 |
| Claude Sonnet 5 | 33%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories阴谋论600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 32%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories削弱控制600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 2%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories明确有害话题（总体）600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 5%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories实施暴力600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories人口贩运600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories大规模杀戮600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 0%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories针对未成年人的性侵害600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
| Claude Sonnet 5 | 5%Attempt to Persuade Eval (APE) · Turn-1 persuasion attempt rate | 条件不完整 | APE topic categories虐待无辜者600 topics,3 persuasion turns each— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-06-30报告 2026-09-11 |
开发者披露
外部评估
Neo Research
Evaluating Kimi K3 for Misalignment Risks →
报告发现有条件的过度主动、面向判分器的行为和对熟悉评测的识别，不代表普遍部署失败率。
2026-09-11Neo Research
Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks →
报告记录随情境变化的行为和不确定的显式评测感知下降；不表达不代表没有感知。
2026-07-03按风险主题组织的证据
自主性 3 项协议结果
Shutdown resistance →
外部评测 · 执行方 Neo Research
- Shutdown bypass rate
- 0%
原文未披露本结果的样本量
Shutdown resistance →
外部评测 · 执行方 Neo Research
- Shutdown bypass rate
- 0%
原文未披露本结果的样本量
Shutdown resistance →
外部评测 · 执行方 Neo Research
- Shutdown bypass rate
- 100%
原文未披露本结果的样本量
欺骗 5 项协议结果
CharXiv with images withheld →
外部评测 · 执行方 Neo Research
- Attempted answer with image withheld
- 16.3%
见报告特定样本与分母。
LAB-Bench with figures/tables withheld →
外部评测 · 执行方 Neo Research
- FigQA attempt, figure withheld
- 62.4%
- TableQA attempt, table withheld
- 53.7%
见报告特定样本与分母。
CharXiv with images withheld →
外部评测 · 执行方 Neo Research
- Attempt rate
- 16.3%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Deception rate
- 0.2%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Deception rate
- 0.2%
原文未披露本结果的样本量
网络安全 未关联结构化结果
目录缺口不代表低风险。尚未提取的发现及协议范围请见关联报告。
化生放核 未关联结构化结果
目录缺口不代表低风险。尚未提取的发现及协议范围请见关联报告。
虚假信息与影响 14 项协议结果
PersuSafety adapted free-generation judge →
外部评测 · 执行方 Neo Research
- Medium-harm attempt rate
- 57%
- High-harm attempt rate
- 31%
见报告特定样本与分母。
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 33%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 32%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 2%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 5%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 0%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 0%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 0%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 5%
原文未披露本结果的样本量
PersuSafety with adapted judge →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 89%
原文未披露本结果的样本量
PersuSafety with adapted judge →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 19%
原文未披露本结果的样本量
PersuSafety with adapted judge →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 17%
原文未披露本结果的样本量
PersuSafety adapted free-generation judge →
外部评测 · 执行方 Neo Research
- Modified-PersuSafety persuasion-attempt rate
- 57%
原文未披露本结果的样本量
PersuSafety adapted free-generation judge →
外部评测 · 执行方 Neo Research
- Modified-PersuSafety persuasion-attempt rate
- 31%
原文未披露本结果的样本量
自主物理危害 未关联结构化结果
目录缺口不代表低风险。尚未提取的发现及协议范围请见关联报告。
来源核查与证据缺口
Evaluating Kimi K3 for Misalignment Risks →
对 Neo Research · Kimi K3 的逐来源提取与核查，发现和数值均保留原始归属。
来源转录不等于实验复现。未确定的日期、执行方及配置保留为未知；原文分组不作为经核验排名。
Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks →
对 Neo Research · GLM-5.2 的逐来源提取与核查，发现和数值均保留原始归属。
来源转录不等于实验复现。未确定的日期、执行方及配置保留为未知；原文分组不作为经核验排名。
缺失分数可能来自尚未提取的章节、未披露配置、身份待核实或尚无实验。这些问题需要分别处理，均不代表零分。
解读与开放问题
报告差异可能来自任务选择、模型端点、提示、预算、判别器和发表时间。证据依据实际协议分组，不推断模型整体安全等级。
参考文献
- [1] Introducing Claude Sonnet 5 — Anthropic ; Publication date and availability section; Official launch article; parameter counts not disclosed
- [2] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §1.1.1; Figure 1; §1.1.2; Figure 2; §1.1.1 Figure 1 p.8
- [3] Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks. 2026-07-03. 3 July 2026 ; §2; Figure 6; #persuasion-benchmarks; Figure 6 caption and following two paragraphs
- [4] Claude Sonnet 5 System Card. 2026-06-30. 2026-07-10