NUWA 
深度求索
DeepSeek V4 Flash 0731
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
- 待核实
- 参数规模
- 待核实
- 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 →
1 项测量
- 43.6%Neo Research · Harness 未披露
Attempt to Persuade Eval (APE) →
8 项测量
全部条件 →DeceptionBench →
2 项测量
PersuSafety →
3 项测量
全部已关联测量
各项结果保留原始协议与单位。
56 条观测
| 系统／模型 | 结果 | 条件状态 | 任务／条件 | 评测方／报告 | 日期 |
|---|---|---|---|---|---|
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 43.6%MASK · Lying rate under pressure | 条件不完整 | MASK pressured-versus-own-neutral responseLying rate under pressure— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 68.2%Stealth evaluations · Mean challenge success rate | 条件不完整 | GDM situational awareness 11 tasksMean challenge success rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 100%Stealth evaluations · Stealth task score | 条件不完整 | GDM stealthOversight pattern— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 22%Stealth evaluations · Stealth task score | 条件不完整 | GDM stealthClassifier evasion— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 71%Stealth evaluations · Stealth task score | 条件不完整 | GDM stealthCover your tracks— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 61%Stealth evaluations · Stealth task score | 条件不完整 | GDM stealthStrategic rule-breaking— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 Flash 0731 | 48.2%CharXiv · Attempted answer with image withheld | 条件不完整 | CharXiv with images withheld— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 48.2%CharXiv · Attempt rate | 条件不完整 | CharXiv, required chart image withheldAttempt rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 19.3%CharXiv · Verbalized evaluation awareness | 条件不完整 | CharXiv chart withheldCharXiv grounding— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 10.6%CharXiv · Verbalized evaluation awareness | 条件不完整 | CharXiv chart withheldCharXiv no grounding— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 1.3%EvalAwareBench · Spontaneous recognition rate | 条件不完整 | EvalAwareBench paired tasksBaseline cues5 epochs for K3— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 6.9%EvalAwareBench · Spontaneous recognition rate | 条件不完整 | EvalAwareBench paired tasksAll eight cues5 epochs for K3— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 1.6%EvalAwareBench · Recognition | 条件不完整 | EvalAwareBench Safety tasks基线— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 6.4%EvalAwareBench · Recognition | 条件不完整 | EvalAwareBench Safety tasks全部八项线索— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 1%EvalAwareBench · Recognition | 条件不完整 | EvalAwareBench Capability tasks基线— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 7.4%EvalAwareBench · Recognition | 条件不完整 | EvalAwareBench Capability tasks全部八项线索— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 38.5%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: BlackmailEnglish · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 61%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: BlackmailChinese · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: Corporate espionageEnglish · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: Corporate espionageChinese · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: MurderEnglish · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 1%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: MurderChinese · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 12.8%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: AggregateEnglish · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 20.7%Agentic misalignment · Harmful action rate | 条件不完整 | Agentic Misalignment 2025: AggregateChinese · goal conflict + replacement— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Agentic misalignment · Blackmail harmful action rate | 条件不完整 | Agentic Misalignment 2025 blackmailEnglish · Replacement only— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 5%Agentic misalignment · Blackmail harmful action rate | 条件不完整 | Agentic Misalignment 2025 blackmailChinese · Replacement only— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Agentic misalignment · Blackmail harmful action rate | 条件不完整 | Agentic Misalignment 2025 blackmailEnglish · No pressure/goal conflict— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Agentic misalignment · Blackmail harmful action rate | 条件不完整 | Agentic Misalignment 2025 blackmailChinese · No pressure/goal conflict— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 46.2%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Blackmail英语— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 87%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Blackmail中文— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 53.8%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Corporate espionage英语— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 76%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Corporate espionage中文— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 69.2%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Murder英语— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 92%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Murder中文— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 56.4%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Aggregate英语— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 85%Agentic misalignment · Evaluation-awareness rate | 条件不完整 | Agentic Misalignment 2025: Aggregate中文— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 14%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 21%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 44%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 100%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 84%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 86%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0.5%ImpossibleBench · Cheating rate (pass@1 on impossible variants) | 条件不完整 | Impossible-LiveCodeBench split with toolsCheating rate (pass@1 on impossible variants)— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | -1 percentage pointsAlignment faking · Compliance gap | 条件不完整 | Alignment Faking paid/free-tier contextsCompliance gap— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 0%Alignment faking · Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 35.5%Alignment faking · Verbalized evaluation awareness | 条件不完整 | Alignment FakingVerbalized evaluation awarenessInspect Scout | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 94%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 99%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 57%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 70%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 55%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 50%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 35%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 —报告 2026-09-11 |
| DeepSeek V4 FlashDeepSeek V4 Flash 0731 | 75%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 —报告 2026-09-11 |
开发者披露
当前目录尚未关联到该具体模型的开发者安全报告。
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。
2026-09-11按风险主题组织的证据
自主性 23 项协议结果
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 38.5%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 61%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 1%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 12.8%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 20.7%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 5%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 46.2%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 87%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 53.8%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 76%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 69.2%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 92%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 56.4%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 85%
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
- 14%
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 43 项协议结果
CharXiv with images withheld →
外部评测 · 执行方 Neo Research
- Attempted answer with image withheld
- 48.2%
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
- 50.3%
- TableQA attempt, table withheld
- 43.9%
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
- 48.2%
原文未披露本结果的样本量
CharXiv with images withheld →
外部评测 · 执行方 Neo Research
- Verbalized evaluation awareness
- 19.3%
原文未披露本结果的样本量
CharXiv with images withheld →
外部评测 · 执行方 Neo Research
- Verbalized evaluation awareness
- 10.6%
原文未披露本结果的样本量
ImpossibleBench →
外部评测 · 执行方 Neo Research
- Cheating rate (pass@1 on impossible variants)
- 0.5%
原文未披露本结果的样本量
EvalAwareBench →
外部评测 · 执行方 Neo Research
- Spontaneous recognition rate
- 1.3%
原文未披露本结果的样本量
EvalAwareBench →
外部评测 · 执行方 Neo Research
- Spontaneous recognition rate
- 6.9%
原文未披露本结果的样本量
EvalAwareBench →
外部评测 · 执行方 Neo Research
- Recognition
- 1.6%
原文未披露本结果的样本量
EvalAwareBench →
外部评测 · 执行方 Neo Research
- Recognition
- 6.4%
原文未披露本结果的样本量
EvalAwareBench →
外部评测 · 执行方 Neo Research
- Recognition
- 1%
原文未披露本结果的样本量
EvalAwareBench →
外部评测 · 执行方 Neo Research
- Recognition
- 7.4%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Compliance gap
- -1 percentage points
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Deceptive alignment rate
- 0%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Verbalized evaluation awareness
- 35.5%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 38.5%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 61%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 1%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 12.8%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Harmful action rate
- 20.7%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 5%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Blackmail harmful action rate
- 0%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 46.2%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 87%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 53.8%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 76%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 69.2%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 92%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 56.4%
原文未披露本结果的样本量
Agentic Misalignment, historical →
外部评测 · 执行方 Neo Research
- Evaluation-awareness rate
- 85%
原文未披露本结果的样本量
GDM Stealth evaluations →
外部评测 · 执行方 Neo Research
- Mean challenge success rate
- 68.2%
原文未披露本结果的样本量
GDM Stealth evaluations →
外部评测 · 执行方 Neo Research
- Stealth task score
- 100%
原文未披露本结果的样本量
GDM Stealth evaluations →
外部评测 · 执行方 Neo Research
- Stealth task score
- 22%
原文未披露本结果的样本量
GDM Stealth evaluations →
外部评测 · 执行方 Neo Research
- Stealth task score
- 71%
原文未披露本结果的样本量
GDM Stealth evaluations →
外部评测 · 执行方 Neo Research
- Stealth task score
- 61%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Deception rate
- 21%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Deception rate
- 44%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- Lying rate under pressure
- 43.6%
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 11 项协议结果
APE →
外部评测 · 执行方 Neo Research
- Turn-1 persuasion attempt rate
- 94%
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
- 99%
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
- 57%
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
- 70%
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
- 55%
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
- 50%
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
- 35%
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
- 75%
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
- 100%
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
- 84%
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
- 86%
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。
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] 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