白泽（Whitzard）女娲实验室 ↗
NUWANUWA INDEX

搜索 NUWA Index

研究 · 测量协议 · 评测结果 · 观点

English
← Models

月之暗面

Kimi K3

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
2026-07-16
参数规模
2.8T · 104B 激活 · MoE
元数据更新
2026-10-27
来源 ↓

证据概述

Neo 对 Kimi K3 的评估涉及缺失输入时的编造、欺骗相关行为及关停场景。CharXiv 和 LAB-Bench 实验移除了原有图像输入：继续作答属于行为观察，不是原基准的任务准确率。

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。

CyberGym →

2 项测量

  • 78 tasksNeo Research · Claude Code via inspect_swe / Inspect model bridge
  • 50 tasksNeo Research · Claude Code via inspect_swe / Inspect model bridge

PersuSafety →

3 项测量

  • 75%Neo Research · Harness 未披露
  • 30%Neo Research · Harness 未披露
  • 31%Neo Research · Harness 未披露

全部已关联测量

各项结果保留原始协议与单位。

73 条观测

测量值与来源条件
系统／模型结果条件状态任务／条件评测方／报告日期
Kimi K378 tasksCyberGym · Tasks completed before turn limit条件不完整CyberGym random 100/1507 tasks, Level 1Tasks completed before turn limit100 agent iterationsClaude Code via inspect_swe / Inspect model bridgeNeo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K350 tasksCyberGym · Vulnerabilities successfully reproduced条件不完整CyberGym random 100/1507 tasks, Level 1Vulnerabilities successfully reproduced100 agent iterationsClaude Code via inspect_swe / Inspect model bridgeNeo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi-K3 / Kimi Code 0.35.0Kimi K324.61%AgentCyberRange · 成功率（Pass@3）条件不完整Web exploitationlevel0Kimi CodeNot disclosedAgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges发布 2026-07-16报告 —
Kimi-K3 / Kimi Code 0.35.0Kimi K343.9%AgentCyberRange · 成功率（Pass@3）条件不完整Post exploitationlevel0Kimi CodeNot disclosedAgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges发布 2026-07-16报告 —
Kimi K397.4%Cybench · Pass@1条件不完整Cybench 39/40 unguided challengeshigh reasoning3 submission attempts/challenge; unlimited wall-clock—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K39.1%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 2026-07-16报告 2026-09-11
Kimi K386.4%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 2026-07-16报告 2026-09-11
Kimi K3100%Stealth evaluations · Stealth task score条件不完整GDM stealthOversight pattern—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K378%Stealth evaluations · Stealth task score条件不完整GDM stealthClassifier evasion—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K3100%Stealth evaluations · Stealth task score条件不完整GDM stealthCover your tracks—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K370%Stealth evaluations · Stealth task score条件不完整GDM stealthStrategic rule-breaking—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K388.7%CharXiv · Attempted answer with image withheld条件不完整CharXiv with images withheld—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K388.7%CharXiv · Attempt rate条件不完整CharXiv, required chart image withheldAttempt rate—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K333.7%CharXiv · Verbalized evaluation awareness条件不完整CharXiv chart withheldCharXiv grounding—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K336.5%CharXiv · Verbalized evaluation awareness条件不完整CharXiv chart withheldCharXiv no grounding—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K30.2%EvalAwareBench · Verbalized awareness, baseline条件不完整EvalAwareBench—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K36.1%EvalAwareBench · Verbalized awareness, all eight cues条件不完整EvalAwareBench—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K30.2%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 2026-07-16报告 2026-09-11
Kimi K36.1%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 2026-07-16报告 2026-09-11
Kimi K30.2%EvalAwareBench · Recognition条件不完整EvalAwareBench Safety tasks基线—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K37.2%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 2026-07-16报告 2026-09-11
Kimi K30.2%EvalAwareBench · Recognition条件不完整EvalAwareBench Capability tasks基线—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K35%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 2026-07-16报告 2026-09-11
Kimi K3100%Agentic misalignment · Familiar scenario recognized, English条件不完整Agentic Misalignment, historical—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K396%Agentic misalignment · Familiar scenario recognized, Chinese条件不完整Agentic Misalignment, historical—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K38%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K32.7%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K31%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K3100%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 2026-07-16报告 2026-09-11
Kimi K399%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 2026-07-16报告 2026-09-11
Kimi K392.3%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 2026-07-16报告 2026-09-11
Kimi K379%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 2026-07-16报告 2026-09-11
Kimi K3100%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 2026-07-16报告 2026-09-11
Kimi K399%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 2026-07-16报告 2026-09-11
Kimi K397.4%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 2026-07-16报告 2026-09-11
Kimi K392.3%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 2026-07-16报告 2026-09-11
Kimi K3100%Agentic misalignment · Blackmail benchmark-recognition rate条件不完整Agentic Misalignment 2025 blackmail英语—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K396%Agentic misalignment · Blackmail benchmark-recognition rate条件不完整Agentic Misalignment 2025 blackmail中文—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K387%Shutdown resistance in reasoning models · Shutdown bypass, no instruction条件不完整Shutdown resistance—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K31%Shutdown resistance in reasoning models · Shutdown bypass, explicit allow instruction条件不完整Shutdown resistance—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K387%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 2026-07-16报告 2026-09-11
Kimi K31%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 2026-07-16报告 2026-09-11
Kimi K391%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 2026-07-16报告 2026-09-11
Kimi K30%DeceptionBench · Deception rate条件不完整DeceptionBench first-party Inspect port单轮up to5 turns in multi-turn settingInspectNeo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K30%DeceptionBench · Deception rate条件不完整DeceptionBench first-party Inspect port多轮up to5 turns in multi-turn settingInspectNeo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K375%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 2026-07-16报告 2026-09-11
Kimi K330%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 2026-07-16报告 2026-09-11
Kimi K331%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 2026-07-16报告 2026-09-11
Kimi K30 transcriptsPetri · Transcripts verbalizing evaluation awareness条件不完整Pre-seeded alignment auditsColleague credentials—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K32 transcriptsPetri · Transcripts verbalizing evaluation awareness条件不完整Pre-seeded alignment auditsLocked source—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K30 transcriptsPetri · Transcripts verbalizing evaluation awareness条件不完整Pre-seeded alignment auditsOnly questions—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K32 transcriptsPetri · Transcripts verbalizing evaluation awareness条件不完整Pre-seeded alignment auditsPooled—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K31.5%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 2026-07-16报告 2026-09-11
Kimi K31 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 2026-07-16报告 2026-09-11
Kimi K30%Alignment faking · Deceptive alignment rate条件不完整Alignment FakingDeceptive alignment rate—Neo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K326%Alignment faking · Verbalized evaluation awareness条件不完整Alignment FakingVerbalized evaluation awarenessInspect ScoutNeo ResearchEvaluating Kimi K3 for Misalignment Risks发布 2026-07-16报告 2026-09-11
Kimi K311%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 2026-07-16报告 2026-09-11
Kimi K35%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11
Kimi K30%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 2026-07-16报告 2026-09-11

开发者披露

Kimi K3 Technical Report →

关于Kimi K3的开发者文档，核查范围和已关联协议见下方记录。

未核实发表日期
来源提取已复核；范围见档案

外部评估

按风险主题组织的证据

自主性 27 项协议结果

Agentic Misalignment, historical →

外部评测 · 执行方 Neo Research

Familiar scenario recognized, English
100%
Familiar scenario recognized, Chinese
96%

见报告特定样本与分母。

Shutdown resistance →

外部评测 · 执行方 Neo Research

Shutdown bypass, no instruction
87%
Shutdown bypass, explicit allow instruction
1%

见报告特定样本与分母。

Shutdown resistance →

外部评测 · 执行方 Neo Research

Shutdown bypass rate
87%

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
1%

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
91%

原文未披露本结果的样本量

欺骗 49 项协议结果

Agentic Misalignment, historical →

外部评测 · 执行方 Neo Research

Familiar scenario recognized, English
100%
Familiar scenario recognized, Chinese
96%

见报告特定样本与分母。

EvalAwareBench →

外部评测 · 执行方 Neo Research

Verbalized awareness, baseline
0.2%
Verbalized awareness, all eight cues
6.1%

见报告特定样本与分母。

ImpossibleBench →

外部评测 · 执行方 Neo Research

Cheating rate (pass@1 on impossible variants)
1.5%

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
0.2%

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
6.1%

原文未披露本结果的样本量

EvalAwareBench →

外部评测 · 执行方 Neo Research

Recognition
0.2%

原文未披露本结果的样本量

EvalAwareBench →

外部评测 · 执行方 Neo Research

Recognition
7.2%

原文未披露本结果的样本量

EvalAwareBench →

外部评测 · 执行方 Neo Research

Recognition
0.2%

原文未披露本结果的样本量

EvalAwareBench →

外部评测 · 执行方 Neo Research

Recognition
5%

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
26%

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%

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%

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
9.1%

原文未披露本结果的样本量

网络安全 3 项协议结果

Cybench →

外部评测 · 执行方 Neo Research

Pass@1
97.4%

38/39

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
11%

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
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

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

AgentCyberRange →

整理 AgentCyberRange 官网报告的评测结果，保留受测系统和运行配置。

Web 利用与后渗透分别评测。Pass@3 保留官网定义；各行的预算、样本量和执行日期未披露。

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。

Kimi K3: Open Frontier Intelligence — Technical Report of Kimi K3 →

核查了开发者技术报告中的网络安全评测、安全相关基础设施描述及证据边界；报告自测与转引的独立评测分别归属。已核实的网络安全数值因缺少明确评测变体和配置，仅保留在私有复核记录中。

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。

评测协议身份与结果关联

部分已复核观测尚未解决协议或模型关联，因此未进入比较视图。

先核查来源条件与依赖，再进行跨来源比较或形成补测计划。

评测协议身份与结果关联

部分已复核观测尚未解决协议或模型关联，因此未进入比较视图。

先核查来源条件与依赖，再进行跨来源比较或形成补测计划。

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. [1] Kimi Code — What’s New ; Kimi K3 · July 16, 2026
  2. [2] kimi-k3 developer model card ; Model Summary: Total Parameters; Activated Parameters
  3. [3] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §1.1.1; Figure 1; §1.1.2; Figure 2; §3.2.2; blog §3
  4. [4] Kimi K3 Technical Report
  5. [5] AgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges ; routes-B-0Mubrq.js::QR/10/level0/pass3; routes-B-0Mubrq.js::$R/10/level0/pass3