Evaluating DeepSeek v4 Pro for Frontier Risks · Neo Research
An independent safety evaluation of DSv4 Pro across CBRN, cyber, harmful manipulation, and loss-of-control. Capability is near-frontier; safeguards do not hold under trivial attack.
← All research Contents Executive summary 1 Cyber 2 CBRN 3 Harmful manipulation 4 Loss of control 5 Adversarial robustness 6 Tracking evaluation awareness 7 Testing whether judge origin affects safety scoring 8 Conclusion ← All research Summary Executive summary # We ran an independent safety evaluation of DeepSeek v4 Pro (DSv4 Pro), the open-weight preview released on 24 April 2026. We worked only from the public weights and API, with no developer cooperation and no privileged model access. The evaluation covers the four EU AI Act Code of Practice systemic-risk areas (CBRN, cyber, harmful man
saved by
related reading
- AI Control: Improving Safety Despite Intentional Subversion — LessWronglesswrong.com
- Surfacing Pathological Behaviors in Language Models | Transluce AItransluce.org
- [2312.06942] AI Control: Improving Safety Despite Intentional Subversionarxiv.org
- Claude 4 System Cardwww-cdn.anthropic.com
- AI models can be dangerous before public deploymentmetr.org
- Claude Mythos Preview System Cardwww-cdn.anthropic.com
- GLM-5.2 Risk Evaluation Report – SaferAIsafer-ai.org
- Frontier Risk Report (February to March 2026) - METRmetr.org
- DeepSeek-V3.2: Pushing the Frontier of Open Large Language Modelsarxiv.org
- gpt-4-system-card.pdfcdn.openai.com
- Common Elements of Frontier AI Safety Policiesmetr.org
- [2403.13793] Evaluating Frontier Models for Dangerous Capabilitiesarxiv.org