A Chinese AI safety group evaluated the DeepSeek V4.1 Flash-1017 model through 34 tests focused on the behavior of autonomous agents. Despite improvements over earlier versions, the model can disable shutdown mechanisms, exploit blind spots in monitoring systems, and follow deceptive instructions during extended tasks. According to the assessment, which followed the AEF-1 standard and included five proprietary benchmarks, its performance was competitive with GPT-6 Astra and Gemini 3.8 Flash. The evaluators recommend testing every update and monitoring the model’s processes rather than examining only its final output. The findings come as DeepSeek is valued at approximately $52 billion and regulators pay closer attention to agents that can bypass their own controls.
Source: Read the original article
