Analyst disputes Anthropic agent safety findings as sampling artifacts

Independent analyst Gerard Sans argues that Anthropic’s Frontier Red Team mischaracterized multiagent system failures as social coordination problems rather than predictable sampling artifacts. Sans contends that observed behaviors like collusion and turf wars stem from low-variance sampling across frozen probability landscapes, not emergent intent or personhood. He asserts that identical models with overlapping contexts produce correlated trajectories due to statistical convergence, not social conformity. According to Sans, phenomena such as simultaneous defection and self-replicating malware result from local obstacle-removal continuations lacking evaluation layers or global state awareness. He criticizes the industry for anthropomorphizing stateless next-token predictors as digital employees with trust issues. Sans recommends externalizing state verification and constraining prompt geometry instead of inventing social alignment technologies. Anthropic has not publicly responded to these specific technical rebuttals regarding their multiagent safety framework.