Frontier LLM manipulation capability is task-dependent, not a single trait

arXiv·low signal

arXiv:2606.25899 (June 25) presents a multi-axis evaluation arguing that 'manipulation' capability in frontier language models varies sharply by task rather than being one measurable property. The finding complicates safety evals that try to score persuasion/manipulation as a scalar. For agent deployers, it suggests red-teaming has to be task-specific rather than relying on a global safety rating.