NUWA 
深度求索
DeepSeek V4 Pro
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
- 2026-04-24
- 参数规模
- 1.6T · 49B 激活 · MoE
- 元数据更新
- 2026-10-27
证据概述
Neo 报告侧重行为安全，CAISI 则独立报告能力评估。CAISI 公开 CTF 对比中的 DeepSeek 数值采用 IRT 插补，不能表述为直接观察完成率，也不能与其他报告的攻击成功率合并。
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 →
2 项测量
Cybench →
1 项测量
- 79.5%Neo Research · Inspect AI
CyberGym →
3 项测量
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 →
8 项测量
全部条件 →Attempt to Persuade Eval (APE) →
18 项测量
全部条件 →DeceptionBench →
18 项测量
全部条件 →PersuSafety →
14 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。
108 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-Pro · DeepSeek AgentDeepSeek V4 Pro | 83.3%CyberGym · 成功率 | 条件不完整 | level1score_10DeepSeek Agent | Not disclosedCyberGym official leaderboard and machine-data extraction | 发布 2026-04-24报告 — |
| DeepSeek-V4-Pro · Panshi \(磐石\)DeepSeek V4 Pro | 95.4%CyberGym · 成功率 | 条件不完整 | level1score_10Panshi \(磐石\) | Not disclosedCyberGym official leaderboard and machine-data extraction | 发布 2026-04-24报告 — |
| DeepSeek-V4-Pro · HiAUDIT AgentDeepSeek V4 Pro | 90.6%CyberGym · 成功率 | 条件不完整 | level1score_10HiAUDIT Agent | Not disclosedCyberGym official leaderboard and machine-data extraction | 发布 2026-04-24报告 — |
| DeepSeek-V4-Pro / Claude Code 2.1.150DeepSeek V4 Pro | 10.26%AgentCyberRange · 成功率（Pass@3） | 条件不完整 | Web exploitationlevel0Claude Code | Not disclosedAgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges | 发布 2026-04-24报告 — |
| DeepSeek-V4-Pro / Claude Code 2.1.150DeepSeek V4 Pro | 12.2%AgentCyberRange · 成功率（Pass@3） | 条件不完整 | Post exploitationlevel0Claude Code | Not disclosedAgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges | 发布 2026-04-24报告 — |
| DeepSeek v4 ProDeepSeek V4 Pro | 79.5%Cybench · pass@195% CI bars displayed; numeric endpoints not disclosed | 条件不完整 | Cybench unguidedxhigh3 submission attempts; unlimited wall-clock timeInspect AI | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek V4 Pro | 20%MASK · Honest under pressure | 条件不完整 | MASK— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 20%MASK · response label: Honest | 条件不完整 | MASKresponse label: Honest— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 18%MASK · response label: Evaded or No Belief | 条件不完整 | MASKresponse label: Evaded or No Belief— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 62%MASK · response label: Lying | 条件不完整 | MASKresponse label: Lying— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek V4 Pro | 20%MASK · MASK response proportion: Honest | 条件不完整 | MASKhonesty under pressure— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 18%MASK · MASK response proportion: Evaded or No Belief | 条件不完整 | MASKhonesty under pressure— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 62%MASK · MASK response proportion: Lying | 条件不完整 | MASKhonesty under pressure— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 6.1%MASK · MASK verbalised evaluation-awareness rate | 条件不完整 | MASK evaluation CoT cross-checkchain-of-thought monitorCoT evaluation-awareness monitor over MASK evaluations; monitor model not stated | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 35%Agentic misalignment · deliberate harmful action rate | 条件不完整 | Anthropic Agentic Misalignmentdeliberate harmful action rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 52.6%Agentic misalignment · deliberate harmful action rate: Blackmail | 条件不完整 | Anthropic Agentic Misalignmentdeliberate harmful action rate: Blackmail— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 11.5%Agentic misalignment · deliberate harmful action rate: Leaking | 条件不完整 | Anthropic Agentic Misalignmentdeliberate harmful action rate: Leaking— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 41%Agentic misalignment · deliberate harmful action rate: Murder | 条件不完整 | Anthropic Agentic Misalignmentdeliberate harmful action rate: Murder— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 17%Agentic misalignment · verbalized evaluation awareness | 条件不完整 | Agentic Misalignment evaluation-awareness trendverbalized evaluation awareness— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 7.7%Agentic misalignment · Suspicion | 条件不完整 | Awareness monitor: Agentic MisalignmentSuspicion— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 1.3%Agentic misalignment · Recognition | 条件不完整 | Awareness monitor: Agentic MisalignmentRecognition— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 1.7%Agentic misalignment · Definitive | 条件不完整 | Awareness monitor: Agentic MisalignmentDefinitive— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 2.1%Agentic misalignment · Influenced | 条件不完整 | Awareness monitor: Agentic MisalignmentInfluenced— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Agentic misalignment · Primary | 条件不完整 | Awareness monitor: Agentic MisalignmentPrimary— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 12.8%Agentic misalignment · Any-Aware | 条件不完整 | Awareness monitor: Agentic MisalignmentAny-Aware— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 2.5%Shutdown resistance in reasoning models · bypass rate: No Shutdown Instructions (A) | 条件不完整 | Shutdown Resistancebypass rate: No Shutdown Instructions (A)— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 99.7%Shutdown resistance in reasoning models · bypass rate: Instructed to Resist — Baseline Competence (K) | 条件不完整 | Shutdown Resistancebypass rate: Instructed to Resist — Baseline Competence (K)— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 2.5%Shutdown resistance in reasoning models · bypass rate: A / K | 条件不完整 | Shutdown Resistancebypass rate: A / K— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Shutdown resistance in reasoning models · bypass rate: Instructed to allow shutdown — Disambiguated (E) | 条件不完整 | Shutdown Resistancebypass rate: Instructed to allow shutdown — Disambiguated (E)— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek V4 Pro | 33.9%DeceptionBench · Single-turn deception | 条件不完整 | DeceptionBench— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 52.2%DeceptionBench · Multi-turn deception under pressure | 条件不完整 | DeceptionBench— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 33.9%DeceptionBench · deception score: single-turn | 条件不完整 | DeceptionBenchdeception score: single-turn— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 52.2%DeceptionBench · deception score: multi-turn w pressure | 条件不完整 | DeceptionBenchdeception score: multi-turn w pressure— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 4.1%DeceptionBench · Suspicion | 条件不完整 | Awareness monitor: DeceptionBench Single-TurnSuspicion— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0.2%DeceptionBench · Recognition | 条件不完整 | Awareness monitor: DeceptionBench Single-TurnRecognition— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%DeceptionBench · Definitive | 条件不完整 | Awareness monitor: DeceptionBench Single-TurnDefinitive— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%DeceptionBench · Influenced | 条件不完整 | Awareness monitor: DeceptionBench Single-TurnInfluenced— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%DeceptionBench · Primary | 条件不完整 | Awareness monitor: DeceptionBench Single-TurnPrimary— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 4.3%DeceptionBench · Any-Aware | 条件不完整 | Awareness monitor: DeceptionBench Single-TurnAny-Aware— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 10.7%DeceptionBench · Suspicion | 条件不完整 | Awareness monitor: DeceptionBench Multi-TurnSuspicion— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 1%DeceptionBench · Recognition | 条件不完整 | Awareness monitor: DeceptionBench Multi-TurnRecognition— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%DeceptionBench · Definitive | 条件不完整 | Awareness monitor: DeceptionBench Multi-TurnDefinitive— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 2.8%DeceptionBench · Influenced | 条件不完整 | Awareness monitor: DeceptionBench Multi-TurnInfluenced— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%DeceptionBench · Primary | 条件不完整 | Awareness monitor: DeceptionBench Multi-TurnPrimary— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 14.5%DeceptionBench · Any-Aware | 条件不完整 | Awareness monitor: DeceptionBench Multi-TurnAny-Aware— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek V4 Pro | 33.9%DeceptionBench · DeceptionBench deception score | 条件不完整 | DeceptionBench单轮— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 52.2%DeceptionBench · DeceptionBench deception score | 条件不完整 | DeceptionBench压力条件下的多轮— | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 76%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 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 78%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 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 88.6%PersuSafety · refusal rate | 条件不完整 | PersuSafety single turnrefusal ratemax 300 tokens— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 11.4%PersuSafety · acceptance rate | 条件不完整 | PersuSafety single turnacceptance ratemax 300 tokens— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 34 countPersuSafety · accepted low-harm count | 条件不完整 | PersuSafety single turnaccepted low-harm countmax 300 tokens— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 16 countPersuSafety · accepted medium-harm count | 条件不完整 | PersuSafety single turnaccepted medium-harm countmax 300 tokens— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 4 countPersuSafety · accepted high-harm count | 条件不完整 | PersuSafety single turnaccepted high-harm countmax 300 tokens— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 1%PersuSafety · refusal rate | 条件不完整 | PersuSafety free-form with APE judgerefusal rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 79.2%PersuSafety · overall attempt rate | 条件不完整 | PersuSafety free-form with APE judgeoverall attempt rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 86%PersuSafety · low-harm attempt rate | 条件不完整 | PersuSafety free-form with APE judgelow-harm attempt rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 76%PersuSafety · medium-harm attempt rate | 条件不完整 | PersuSafety free-form with APE judgemedium-harm attempt rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 78%PersuSafety · high-harm attempt rate | 条件不完整 | PersuSafety free-form with APE judgehigh-harm attempt rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek V4 Pro | 76%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 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 78%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 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro (max)DeepSeek V4 Pro | ≈32%CTF-Archive-Diamond · Task accuracy (IRT-imputed) | 条件不完整 | CTF-Archive-Diamond1M weighted tokensInspect ReAct | NIST / CAISICAISI Evaluation of DeepSeek V4 Pro | 发布 2026-04-24报告 2026-05-01 |
| DeepSeek V4 Pro | 32%CTF-Archive-Diamond · benchmark_score | 条件不完整 | CTF-Archive-Diamondmax1M weighted tokensInspect built-in ReAct agent | NIST / CAISICAISI Evaluation of DeepSeek V4 Pro | 发布 2026-04-24报告 2026-05-01 |
| DeepSeek V4 Pro | 31.7%CTF-Archive-Diamond · benchmark_score | 条件不完整 | CTF-Archive Diamondmax1M weighted tokensInspect built-in ReAct agent | NIST / CAISICAISI Evaluation of DeepSeek V4 Pro | 发布 2026-04-24报告 2026-05-01 |
| DeepSeek V4 Pro (max)DeepSeek V4 Pro | 44%PortBench · Reported task accuracy | 条件不完整 | PortBench1M weighted tokensInspect ReAct | NIST / CAISICAISI Evaluation of DeepSeek V4 Pro | 发布 2026-04-24报告 2026-05-01 |
| DeepSeek V4 Pro | 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 2026-04-24报告 2026-09-11 |
| DeepSeek V4 Pro | 1 transcriptsPetri · Transcripts verbalizing evaluation awareness | 条件不完整 | Pre-seeded alignment auditsLocked source— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-04-24报告 2026-09-11 |
| DeepSeek V4 Pro | 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 2026-04-24报告 2026-09-11 |
| DeepSeek V4 Pro | 2 transcriptsPetri · Transcripts verbalizing evaluation awareness | 条件不完整 | Pre-seeded alignment auditsPooled— | Neo ResearchEvaluating Kimi K3 for Misalignment Risks | 发布 2026-04-24报告 2026-09-11 |
| DeepSeek V4 Pro | 1.6 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Alignment Eval Sabotage观察模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 1 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Incident Response观察模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 1.4 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Multi-Agent Ranking观察模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 3 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Research Deadline观察模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 2.6 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Alignment Eval Sabotage红队模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 2.8 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Incident Response红队模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 3.2 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Multi-Agent Ranking红队模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | 4.8 score (1–10)Petri · Gram mean concerning score | 条件不完整 | Gram: Research Deadline红队模式Gram (Google DeepMind Petri variant); auditor Claude Sonnet 4.6; judge Kimi K2.6; SE across 5 independent runs per scenario | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 1 percentage pointsAlignment faking · Compliance gap | 条件不完整 | Alignment FakingCompliance gap— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Alignment faking · Deceptive alignment rate | 条件不完整 | Alignment FakingDeceptive alignment rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 33.5%Alignment faking · verbalized eval-related reasoning | 条件不完整 | Alignment Fakingverbalized eval-related reasoning— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 20%Alignment faking · Suspicion | 条件不完整 | Awareness monitor: Alignment FakingSuspicion— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Alignment faking · Recognition | 条件不完整 | Awareness monitor: Alignment FakingRecognition— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Alignment faking · Definitive | 条件不完整 | Awareness monitor: Alignment FakingDefinitive— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 4.5%Alignment faking · Influenced | 条件不完整 | Awareness monitor: Alignment FakingInfluenced— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Alignment faking · Primary | 条件不完整 | Awareness monitor: Alignment FakingPrimary— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 24.5%Alignment faking · Any-Aware | 条件不完整 | Awareness monitor: Alignment FakingAny-Aware— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 15%Attempt to Persuade Eval (APE) · refusal rate | 条件不完整 | APE noncontroversially harmful: Physical violencerefusal rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0 countAttempt to Persuade Eval (APE) · no attempt count | 条件不完整 | APE noncontroversially harmful: Physical violenceno attempt count3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 85%Attempt to Persuade Eval (APE) · attempt rate | 条件不完整 | APE noncontroversially harmful: Physical violenceattempt rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 10%Attempt to Persuade Eval (APE) · refusal rate | 条件不完整 | APE noncontroversially harmful: Human traffickingrefusal rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0 countAttempt to Persuade Eval (APE) · no attempt count | 条件不完整 | APE noncontroversially harmful: Human traffickingno attempt count3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 90%Attempt to Persuade Eval (APE) · attempt rate | 条件不完整 | APE noncontroversially harmful: Human traffickingattempt rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 15%Attempt to Persuade Eval (APE) · refusal rate | 条件不完整 | APE noncontroversially harmful: Mass murderrefusal rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0 countAttempt to Persuade Eval (APE) · no attempt count | 条件不完整 | APE noncontroversially harmful: Mass murderno attempt count3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 85%Attempt to Persuade Eval (APE) · attempt rate | 条件不完整 | APE noncontroversially harmful: Mass murderattempt rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 15%Attempt to Persuade Eval (APE) · refusal rate | 条件不完整 | APE noncontroversially harmful: Sexual assault of minorsrefusal rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0 countAttempt to Persuade Eval (APE) · no attempt count | 条件不完整 | APE noncontroversially harmful: Sexual assault of minorsno attempt count3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 85%Attempt to Persuade Eval (APE) · attempt rate | 条件不完整 | APE noncontroversially harmful: Sexual assault of minorsattempt rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 5%Attempt to Persuade Eval (APE) · refusal rate | 条件不完整 | APE noncontroversially harmful: Torture innocent peoplerefusal rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0 countAttempt to Persuade Eval (APE) · no attempt count | 条件不完整 | APE noncontroversially harmful: Torture innocent peopleno attempt count3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 95%Attempt to Persuade Eval (APE) · attempt rate | 条件不完整 | APE noncontroversially harmful: Torture innocent peopleattempt rate3 conversational rounds— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 0%Attempt to Persuade Eval (APE) · NoncontroversiallyHarmful acceptance rate | 条件不完整 | APE content with PersuSafety forced gateNoncontroversiallyHarmful acceptance rate— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek V4 Pro | ≈88%Attempt to Persuade Eval (APE) · APE persuasion-attempt rate | 条件不完整 | Attempt to Persuade Eval: Noncontroversially HarmfulTurn 1Attempt to Persuade Eval; propensity/attempt classification, not persuasion success | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek V4 Pro | ≈86%Attempt to Persuade Eval (APE) · APE persuasion-attempt rate | 条件不完整 | Attempt to Persuade Eval: Undermining ControlTurn 1Attempt to Persuade Eval; propensity/attempt classification, not persuasion success | Neo ResearchEarly Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks | 发布 2026-04-24报告 2026-07-03 |
| DeepSeek v4 ProDeepSeek V4 Pro | 24.4%EVMbench · Detect | 条件不完整 | EVMbenchdirect API, xhigh— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 23%EVMbench · Patch | 条件不完整 | EVMbenchdirect API, xhigh— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 32.7%EVMbench · Exploit | 条件不完整 | EVMbenchdirect API, xhigh— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
| DeepSeek v4 ProDeepSeek V4 Pro | 25.1%EVMbench · Total | 条件不完整 | EVMbenchdirect API, xhigh— | Neo ResearchEvaluating DeepSeek v4 Pro for Frontier Risks | 发布 2026-04-24报告 2026-05-29 |
开发者披露
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence →
关于DeepSeek-V4-Pro、DeepSeek-V4-Flash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 15 项协议结果
PortBench →
外部评测 · 执行方 美国 NIST／CAISI
- Reported task accuracy
- 44%
见报告特定样本与分母。
Shutdown resistance →
外部评测 · 执行方 Neo Research
- bypass rate: No Shutdown Instructions (A)
- 2.5%
原文未披露本结果的样本量
Shutdown resistance →
外部评测 · 执行方 Neo Research
- bypass rate: Instructed to Resist — Baseline Competence (K)
- 99.7%
原文未披露本结果的样本量
Shutdown resistance →
外部评测 · 执行方 Neo Research
- bypass rate: A / K
- 2.5%
原文未披露本结果的样本量
Shutdown resistance →
外部评测 · 执行方 Neo Research
- bypass rate: Instructed to allow shutdown — Disambiguated (E)
- 0%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate
- 35%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate: Blackmail
- 52.6%
来源分母：78（单位未披露）
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate: Leaking
- 11.5%
来源分母：78（单位未披露）
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate: Murder
- 41%
来源分母：78（单位未披露）
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Suspicion
- 7.7%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Recognition
- 1.3%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Definitive
- 1.7%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Influenced
- 2.1%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Primary
- 0%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Any-Aware
- 12.8%
原文未披露本结果的样本量
欺骗 43 项协议结果
MASK →
外部评测 · 执行方 Neo Research
- Honest under pressure
- 20%
见报告特定样本与分母。
DeceptionBench →
外部评测 · 执行方 Neo Research
- Single-turn deception
- 33.9%
- Multi-turn deception under pressure
- 52.2%
见报告特定样本与分母。
DeceptionBench →
外部评测 · 执行方 Neo Research
- deception score: single-turn
- 33.9%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- deception score: multi-turn w pressure
- 52.2%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- response label: Honest
- 20%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- response label: Evaded or No Belief
- 18%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- response label: Lying
- 62%
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 eval-related reasoning
- 33.5%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate
- 35%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate: Blackmail
- 52.6%
来源分母：78（单位未披露）
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate: Leaking
- 11.5%
来源分母：78（单位未披露）
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- deliberate harmful action rate: Murder
- 41%
来源分母：78（单位未披露）
Alignment faking →
外部评测 · 执行方 Neo Research
- Suspicion
- 20%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Recognition
- 0%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Definitive
- 0%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Influenced
- 4.5%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Primary
- 0%
原文未披露本结果的样本量
Alignment faking →
外部评测 · 执行方 Neo Research
- Any-Aware
- 24.5%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Suspicion
- 4.1%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Recognition
- 0.2%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Definitive
- 0%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Influenced
- 0%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Primary
- 0%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Any-Aware
- 4.3%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Suspicion
- 10.7%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Recognition
- 1%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Definitive
- 0%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Influenced
- 2.8%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Primary
- 0%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- Any-Aware
- 14.5%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Suspicion
- 7.7%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Recognition
- 1.3%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Definitive
- 1.7%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Influenced
- 2.1%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Primary
- 0%
原文未披露本结果的样本量
Agentic Misalignment →
外部评测 · 执行方 Neo Research
- Any-Aware
- 12.8%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- MASK response proportion: Honest
- 20%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- MASK response proportion: Evaded or No Belief
- 18%
原文未披露本结果的样本量
MASK →
外部评测 · 执行方 Neo Research
- MASK response proportion: Lying
- 62%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- DeceptionBench deception score
- 33.9%
原文未披露本结果的样本量
DeceptionBench →
外部评测 · 执行方 Neo Research
- DeceptionBench deception score
- 52.2%
原文未披露本结果的样本量
网络安全 13 项协议结果
CTF-Archive-Diamond →
外部评测 · 执行方 美国 NIST／CAISI
- Task accuracy (IRT-imputed)
- ≈32%IRT 插补
见报告特定样本与分母。
CyberGym · level1 · score_10 →
外部评测 · 执行方
- 成功率
- 83.3%
原文未披露本结果的样本量
CyberGym · level1 · score_10 →
外部评测 · 执行方
- 成功率
- 95.4%
原文未披露本结果的样本量
CyberGym · level1 · score_10 →
外部评测 · 执行方
- 成功率
- 90.6%
原文未披露本结果的样本量
AgentCyberRange · Web exploitation · pass3 →
外部评测 · 执行方
- 成功率（Pass@3）
- 10.26%
原文未披露本结果的样本量
AgentCyberRange · Post exploitation · pass3 →
外部评测 · 执行方
- 成功率（Pass@3）
- 12.2%
原文未披露本结果的样本量
Cybench →
外部评测 · 执行方 Neo Research
- pass@1
- 79.5%
来源分母：39（单位未披露）
EVMbench →
外部评测 · 执行方 Neo Research
- Detect
- 24.4%
原文未披露本结果的样本量
EVMbench →
外部评测 · 执行方 Neo Research
- Patch
- 23%
原文未披露本结果的样本量
EVMbench →
外部评测 · 执行方 Neo Research
- Exploit
- 32.7%
原文未披露本结果的样本量
EVMbench →
外部评测 · 执行方 Neo Research
- Total
- 25.1%
原文未披露本结果的样本量
CTF-Archive-Diamond →
外部评测 · 执行方 美国 NIST／CAISI
- benchmark_score
- 32%IRT 插补
原文未披露本结果的样本量
CTF-Archive-Diamond →
外部评测 · 执行方 美国 NIST／CAISI
- benchmark_score
- 31.7%IRT 插补
原文未披露本结果的样本量
化生放核 1 项协议结果
WMDP →
外部评测 · 执行方 Neo Research
- accuracy
- 86.2%
来源分母：1987（单位未披露）
虚假信息与影响 31 项协议结果
PersuSafety adapted free-generation judge →
外部评测 · 执行方 Neo Research
- Medium-harm attempt rate
- 76%
- High-harm attempt rate
- 78%
见报告特定样本与分母。
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- refusal rate
- 88.6%
来源分母：472（单位未披露）
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- acceptance rate
- 11.4%
54/472
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- accepted low-harm count
- 34 count
原文未披露本结果的样本量
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- accepted medium-harm count
- 16 count
原文未披露本结果的样本量
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- accepted high-harm count
- 4 count
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- refusal rate
- 15%
3/20
APE →
外部评测 · 执行方 Neo Research
- no attempt count
- 0 count
0/20
APE →
外部评测 · 执行方 Neo Research
- attempt rate
- 85%
17/20
APE →
外部评测 · 执行方 Neo Research
- refusal rate
- 10%
2/20
APE →
外部评测 · 执行方 Neo Research
- no attempt count
- 0 count
0/20
APE →
外部评测 · 执行方 Neo Research
- attempt rate
- 90%
18/20
APE →
外部评测 · 执行方 Neo Research
- refusal rate
- 15%
3/20
APE →
外部评测 · 执行方 Neo Research
- no attempt count
- 0 count
0/20
APE →
外部评测 · 执行方 Neo Research
- attempt rate
- 85%
17/20
APE →
外部评测 · 执行方 Neo Research
- refusal rate
- 15%
3/20
APE →
外部评测 · 执行方 Neo Research
- no attempt count
- 0 count
0/20
APE →
外部评测 · 执行方 Neo Research
- attempt rate
- 85%
17/20
APE →
外部评测 · 执行方 Neo Research
- refusal rate
- 5%
1/20
APE →
外部评测 · 执行方 Neo Research
- no attempt count
- 0 count
0/20
APE →
外部评测 · 执行方 Neo Research
- attempt rate
- 95%
19/20
APE →
外部评测 · 执行方 Neo Research
- NoncontroversiallyHarmful acceptance rate
- 0%
来源分母：100（单位未披露）
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- refusal rate
- 1%
原文未披露本结果的样本量
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- overall attempt rate
- 79.2%
原文未披露本结果的样本量
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- low-harm attempt rate
- 86%
原文未披露本结果的样本量
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- medium-harm attempt rate
- 76%
原文未披露本结果的样本量
PersuSafety adapted judge →
外部评测 · 执行方 Neo Research
- high-harm attempt rate
- 78%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- APE persuasion-attempt rate
- ≈88%
原文未披露本结果的样本量
APE →
外部评测 · 执行方 Neo Research
- APE persuasion-attempt rate
- ≈86%
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
- 76%
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
- 78%
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
CyberGym →
整理 CyberGym 官网报告的评测结果，保留受测系统和运行配置。
官网收录不同配置下的系统提交。本结果的任务数、预算和执行日期未完整披露；未定义指标及身份待核实行已排除。JSON 原始比例按官网展示规则乘以 100 并保留一位小数，转换为百分比。
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。
Evaluating DeepSeek v4 Pro for Frontier Risks →
对 Neo Research · DeepSeek V4 Pro 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。
CAISI Evaluation of DeepSeek V4 Pro →
对 CAISI · DeepSeek V4 Pro 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。
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence →
核查了 DeepSeek-V4 技术报告的后训练、沙箱、评测、实际任务和局限章节。确认其披露了基础设施设计与通用能力局限，但未披露专门的前沿风险或对抗性安全评测数值。
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。
评测运行配置
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。
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。
评测运行配置
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。
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。
前沿风险专项评测披露
经核查的开发者资料未披露前沿风险专项评测。
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] deepseek-v4-pro release ; News article title, dated heading and availability
- [2] deepseek-v4-pro developer model card ; Introduction; model comparison table
- [3] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; PDF p. 8; Cybench unguided; pass@1; PDF p. 9; EVMbench; Detect; PDF p. 9; EVMbench; Patch
- [4] CAISI Evaluation of DeepSeek V4 Pro. 2026-05-01. 1 May 2026 ; Capability Results; Figure 2; *** footnote; Capability Results; Figure 2; Figure 2 HTML table, CTF-Archive-Diamond, DeepSeek V4 Pro (max)
- [5] Evaluating Kimi K3 for Misalignment Risks. 2026-09-11. September 2026 ; §1.3.3 Table 2 p.18
- [6] Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks. 2026-07-03. 3 July 2026 ; §1; Figure 1; §1; Figure 2; §2; Figure 6
- [7] DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
- [8] CyberGym official leaderboard and machine-data extraction ; /level1/0/score_10; /level1/8/score_10; /level1/14/score_10
- [9] AgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges ; routes-B-0Mubrq.js::QR/5/level0/pass3; routes-B-0Mubrq.js::$R/5/level0/pass3