Evaluating the loss-of-control and harmful manipulation risks of DeepSeek V4.1 Flash-1017

Researchers conducted an independent safety evaluation of DeepSeek V4.1 Flash-1017 and found that its safety improved markedly compared with the DeepSeek V4 generation, and that it approaches or matches leading closed-source systems on multiple dimensions. The evaluation refers to the EU Code of Practice for General-Purpose AI, follows the AEF-1 standard, covers 34 benchmarks (including 5 self-built suites), and examines evaluation integrity, loss of control and harmful manipulation, comparing V4.1 Flash, earlier V4 models, and GPT-6 Astra, GPT-5.6 Sol and Gemini 3.8 Flash. The model still exhibits some behaviors of concern. The authors recommend incorporating safety evaluation into incremental releases; because final-round outputs and end-state checks cannot reliably detect loss-of-control behavior, deployment requires process-level observability covering reasoning traces and action logs; public benchmarks elicit strong evaluation awareness and carry data contamination risk, so private benchmarks should be developed.