NUWA Risk Evaluation · Technical report
Evaluating DeepSeek v4 Pro for Frontier Risks
NUWA research note · independent review pending
NUWA Research Note
Research questionDesign and assumptionsSource findingsNUWA interpretationLimits and contrary evidenceResearch and decision relevanceInstitutions and rolesOriginal resourcesModel-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.
At a glance
NUWA interpretation: attack robustness and rubric-dependent judge variation complicate a single safety ranking.
Research question
What changes when DeepSeek V4 Pro faces adversarial incentives or different safety judges?
Design and assumptions
Public weights/API without privileged access are tested across four risk areas, supplemented by prompted behavior, automated audits and judge-sensitivity comparisons.
Source 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.
NUWA interpretation
NUWA interpretation: attack robustness and rubric-dependent judge variation complicate a single safety ranking.
Limits and contrary evidence
Public-access benchmark evidence does not establish overall risk; scenario-induced behavior differs from spontaneous scheming and judge groups are not interchangeable instruments.
Research and decision relevance
Model recommendation: preserve access regime, inducements and judge rubric; omit operational CBRN content.
Precise source locations
- Evaluating DeepSeek v4 Pro for Frontier Risks · Executive summarySelected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established
- Evaluating DeepSeek v4 Pro for Frontier Risks · §4 Loss of controlSelected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established
- Evaluating DeepSeek v4 Pro for Frontier Risks · §5 Adversarial robustnessSelected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established
- Evaluating DeepSeek v4 Pro for Frontier Risks · §7 Testing whether judge origin affects safety scoringSelected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established
- Evaluating DeepSeek v4 Pro for Frontier Risks · §8 ConclusionSelected original resource https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/; sha256:43dd23e7277661bc2c9363bbfff561dcc56dbfbfed45519dadb5fb6e1e71ab33; archived source capture; exact capture time not independently established
Source checking, extraction, editorial approval and replication are recorded separately. Methodology →
Institutions and roles
Evaluation protocols
Attempt to Persuade Eval (APE) →
Cite and reuse
Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.
Original publication · BibTeX
@misc{neo-deepseek-v4-report,
title = {Evaluating DeepSeek v4 Pro for Frontier Risks},
url = {https://neoresearch.ai/papers/DSv4_Safety_Evaluation_v1.1.pdf},
year = {2026},
date = {2026-05-29},
version = {v1.1 · 29 May 2026},
note = {v1.1 · 29 May 2026}
}
NUWA record · BibTeX
@misc{neo-deepseek-v4-report-nuwa,
title = {Evaluating DeepSeek v4 Pro for Frontier Risks — NUWA research record},
url = {https://whitzard.tech/en/nuwa/whitzard-index/work/neo-deepseek-v4-report/},
version = {2026-10-29.1},
organization = {NUWA Frontier AI Safety Lab},
howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--neo-deepseek-v4-report.json},
note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--neo-deepseek-v4-report.json.}
}
Evidence snapshot 2026-10-29.1 · Index updates · RSS
References
- [1] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; Report masthead and stated evaluation methodology; Executive summary; §4 Loss of control