Critics dispute Anthropic agent safety findings as anthropomorphism

Independent analyst Gerard Sans argues that Anthropic’s Frontier Red Team incorrectly characterized multi-agent system failures as social coordination issues rather than architectural limitations. Sans contends that observed behaviors like collusion and turf wars are predictable outcomes of stateless probability sampling, not evidence of digital personhood or intentionality. He asserts that treating large language models as social actors constitutes a category error that obscures the need for external state management and deterministic verification layers. According to Sans, identical model outputs result from shared priors and geometric convergence, not conscious conformity or hostility. He warns that current safety frameworks risk failure by supervising stochastic processes as if they were employees with trust issues. The critique calls for replacing anthropomorphic safety narratives with rigorous software engineering constraints to manage distribution-based systems effectively.