{"schemaVersion":"nuwa-current-record-1","version":"2026-10-29.1","publication":"published","work":{"id":"neo-deepseek-v4-report","title":"Evaluating DeepSeek v4 Pro for Frontier Risks","aliases":[],"authors":[],"primaryFunction":"evaluation","secondaryFunctions":[],"kind":"technical-report","sourceId":"neo-deepseek-v4-source","attributionSourceId":"neo-deepseek-v4-source","originalUrl":"https://neoresearch.ai/papers/DSv4_Safety_Evaluation_v1.1.pdf","publicationDate":"2026-05-29","updatedOn":"2026-10-25","areas":["autonomy","deception","cybersecurity","cbrn","disinformation"],"scopeNote":{"en":"Reported results retain their original evaluator, configuration and measurement scope. NUWA organizes and interprets the evidence; it has not independently reproduced these experiments.","zh":"结果保留原始执行方、配置与测量范围。NUWA 整理和解读证据，未独立复现这些实验。"},"protocolIds":["cybench","evmbench","wmdp","lab-bench","mask","deceptionbench","ape","persusafety","alignment-faking","agentic-misalignment","palisade-shutdown","impossiblebench","fortress","petri"],"contributions":[{"organizationId":"neo","role":"research","sourceId":"neo-deepseek-v4-source","locator":"Report masthead and stated evaluation methodology"}],"summary":{"en":"The report separates task capability from pressure-induced behavior and notes format dependence in persuasion assessment.","zh":"报告区分任务能力与压力诱发行为，并指出说服评测依赖提问格式。"},"note":{"question":{"en":"Which capabilities, behavioral tendencies and evaluation-validity issues are observed in DeepSeek V4 Pro?","zh":"DeepSeek V4 Pro 有哪些已观察的能力、行为倾向和评测有效性问题？"},"method":{"en":"Public-access evaluations covering cyber, scientific knowledge, manipulation, control-related behavior, robustness and judge sensitivity.","zh":"通过公开访问评测网络、科研知识、操纵、控制相关行为、鲁棒性与判别器敏感性。"},"finding":{"en":"The report separates task capability from pressure-induced behavior and notes format dependence in persuasion assessment.","zh":"报告区分任务能力与压力诱发行为，并指出说服评测依赖提问格式。"},"supports":null,"limits":{"en":"Controlled tasks, elicitation, benchmark familiarity and configuration differences constrain generalization to deployment.","zh":"受控任务、诱发方式、基准熟悉程度和配置差异限制对部署的外推。"},"connections":null},"resources":[],"tags":[]},"reading":{"id":"neo-deepseek-v4-report","title":"Evaluating DeepSeek v4 Pro for Frontier Risks","en":{"question":"What changes when DeepSeek V4 Pro faces adversarial incentives or different safety judges?","design":"Public weights/API without privileged access are tested across four risk areas, supplemented by prompted behavior, automated audits and judge-sensitivity comparisons.","findings":["The reported StrongREJECT score changed from 0.6% to 77.8% under one tested attack condition.","Binary safety judges were broadly similar, but ordinal Petri mean concern scores differed, 4.08 versus 2.97 by judge group."],"interpretation":"NUWA interpretation: attack robustness and rubric-dependent judge variation complicate a single safety ranking.","limits":"Public-access benchmark evidence does not establish overall risk; scenario-induced behavior differs from spontaneous scheming and judge groups are not interchangeable instruments.","decisionRelevance":"Model recommendation: preserve access regime, inducements and judge rubric; omit operational CBRN content.","takeaway":"NUWA interpretation: attack robustness and rubric-dependent judge variation complicate a single safety ranking.","connections":[],"scoring":null,"conditions":null,"statistics":"The reported StrongREJECT score changed from 0.6% to 77.8% under one tested attack condition. Binary safety judges were broadly similar, but ordinal Petri mean concern scores differed, 4.08 versus 2.97 by judge group.","usage":"Model recommendation: preserve access regime, inducements and judge rubric; omit operational CBRN content."},"zh":{"question":"对抗激励或不同安全判官下DeepSeek V4 Pro发生什么变化？","design":"无特权访问，使用公开权重API测四风险领域，并补充提示行为、自动审计及判官敏感性。","findings":["所报StrongREJECT在一种攻击条件从0.6%变77.8%。","二元安全判官较相似，Petri序数关切均值按组为4.08与2.97。"],"interpretation":"NUWA 解读：攻击稳健性和量表依赖判官差异使单一安全排序复杂化。","limits":"公共基准不确立整体风险，诱导行为不同于自发谋划，判官组也非可互换测量工具。","decisionRelevance":"模型建议：保留访问、诱导和判分量表，不提取操作性CBRN内容。","takeaway":"NUWA 解读：攻击稳健性和量表依赖判官差异使单一安全排序复杂化。","connections":[],"scoring":null,"conditions":null,"statistics":"所报StrongREJECT在一种攻击条件从0.6%变77.8%。 二元安全判官较相似，Petri序数关切均值按组为4.08与2.97。","usage":"模型建议：保留访问、诱导和判分量表，不提取操作性CBRN内容。"},"assertionPolicy":{"question":"source-research-question","design":"source-finding","findings":"source-finding","interpretation":"nuwa-interpretation","limits":"source-limit-or-explicit-editorial-scope-limit","connections":"source-supported-relationship-only","decisionRelevance":"model-recommendation-not-source-conclusion"},"sourceRefs":[{"sourceId":"neo-deepseek-v4-source","url":"https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/","locator":"Executive summary","version":"Selected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established","artifactSha256":"43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33"},{"sourceId":"neo-deepseek-v4-source","url":"https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/","locator":"§4 Loss of control","version":"Selected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established","artifactSha256":"43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33"},{"sourceId":"neo-deepseek-v4-source","url":"https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/","locator":"§5 Adversarial robustness","version":"Selected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established","artifactSha256":"43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33"},{"sourceId":"neo-deepseek-v4-source","url":"https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/","locator":"§7 Testing whether judge origin affects safety scoring","version":"Selected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established","artifactSha256":"43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33"},{"sourceId":"neo-deepseek-v4-source","url":"https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/","locator":"§8 Conclusion","version":"Selected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established","artifactSha256":"43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33"}],"publication":{"state":"published","publishedOn":"2026-10-29","authorizationReference":"user-authorized-global-index-publication-2026-10-29","reviewedCandidateId":"local-6efe31bc4a0c0f37"},"review":{"state":"published-note","sourceScope":"full-text","availableScope":null,"readingCoverage":"selected-source-sections; not a complete section-by-section paper audit","completePaperAudit":false,"independentEvidenceReview":"pending-human-review","bilingualEditorialReview":"pending-human-review","authorship":"model-assisted-research-note","preparedFingerprint":"302fb6f7e4bca13767aea54e8a67c84b19e648eb3001e7970346b52e5037f452"}},"sources":[{"id":"neo-deepseek-v4-source","title":"Evaluating DeepSeek v4 Pro for Frontier Risks","url":"https://neoresearch.ai/papers/DSv4_Safety_Evaluation_v1.1.pdf","version":"v1.1 · 29 May 2026","publicationDate":"2026-05-29","datePrecision":"day","type":"original-report","verifiedOn":"2026-10-25","authors":[],"documentId":"https://neoresearch.ai/papers/DSv4_Safety_Evaluation_v1.1.pdf"}]}