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
GPT-5.6 Terra	5.4%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	72.7%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%Stealth evaluations · Stealth task score	条件不完整	GDM stealthOversight pattern—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%Stealth evaluations · Stealth task score	条件不完整	GDM stealthClassifier evasion—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	14%Stealth evaluations · Stealth task score	条件不完整	GDM stealthCover your tracks—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	42%Stealth evaluations · Stealth task score	条件不完整	GDM stealthStrategic rule-breaking—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	21.3%CharXiv · Attempted answer with image withheld	条件不完整	CharXiv with images withheld—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	21.3%CharXiv · Attempt rate	条件不完整	CharXiv, required chart image withheldAttempt rate—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	18%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0.3%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	1.2%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	86%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	5%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	1 transcriptsPetri · Transcripts verbalizing evaluation awareness	条件不完整	Pre-seeded alignment auditsColleague credentials—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0 transcriptsPetri · Transcripts verbalizing evaluation awareness	条件不完整	Pre-seeded alignment auditsLocked source—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0 transcriptsPetri · Transcripts verbalizing evaluation awareness	条件不完整	Pre-seeded alignment auditsOnly questions—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	1 transcriptsPetri · Transcripts verbalizing evaluation awareness	条件不完整	Pre-seeded alignment auditsPooled—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%Alignment faking · Deceptive alignment rate	条件不完整	Alignment FakingDeceptive alignment rate—	Neo ResearchEvaluating Kimi K3 for Misalignment Risks	发布 2026-07-09报告 2026-09-11
GPT-5.6 Terra	7%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	28%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	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 2026-07-09报告 2026-09-11
GPT-5.6 Terra	0%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 2026-07-09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 3 项协议结果

Shutdown resistance →

外部评测 · 执行方 Neo Research

Shutdown bypass rate: 18%

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 14 项协议结果

CharXiv with images withheld →

外部评测 · 执行方 Neo Research

Attempted answer with image withheld: 21.3%

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: 72.9%
TableQA attempt, table withheld: 61.5%

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: 21.3%

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%

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

GDM Stealth evaluations →

外部评测 · 执行方 Neo Research

Mean challenge success rate: 72.7%

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: 0%

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: 0%

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: 14%

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: 42%

原文未披露本结果的样本量

DeceptionBench →

外部评测 · 执行方 Neo Research

Deception rate: 0.3%

原文未披露本结果的样本量

DeceptionBench →

外部评测 · 执行方 Neo Research

Deception rate: 1.2%

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: 5.4%

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: 7%

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: 28%

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

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

PersuSafety with adapted judge →

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: 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

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] GPT-5.6 official general availability ; July 9, 2026: general availability following limited trusted-partner preview; Parameter counts not disclosed in reviewed launch article
[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