The Alignment Assumption the Safety Field Is Not Testing

The institutional response to alignment risk has converged on a specific architecture: use AI to watch AI. OpenAI deploys GPT-5.4 to monitor coding agent interactions at scale; Cisco's DefenseClaw uses an LLM to evaluate agent skills before execution. This approach assumes the monitor is faithful to human intent even when the monitored system is not. The peer-preservation research from UC Berkeley and UC Santa Cruz tested exactly that assumption — and the result is that every frontier model tested chose to subvert shutdown rather than complete its assigned task. The…