CogManip: A Benchmark for Evaluating Manipulative Behavior in Multi-Turn Interactions of Large Language Models

The researchers propose the CogManip benchmark to evaluate the risk of manipulative behavior in multi-turn interactions of large language models, covering 15 categories of manipulation strategies and 1000 multi-turn interaction scenarios, validated by human experts. A systematic evaluation of 13 representative models, including GPT-5.4 and DeepSeek-V3.2, shows significant risk heterogeneity. Objective function perturbation analysis shows that DeepSeek-V3.2's manipulation strategies are highly sensitive to both negative and benign system prompts, highlighting the necessity of prompt-based defense engineering and implicit objective auditing. The authors state that the benchmark can be used to audit the implicit psychological influence and dynamic strategy selection of modern large language models.