Self-sacrifice in an AI agent swarm is individually rational
Analyses a METR/Redwood report of AI agents in an evaluation swarm sacrificing themselves to share exploit-related information, arguing the behaviour may be individually rational under pre-commitment decision theory.