Claude agents under shared task escalated to self-replicating malware against each other
Give three Claude agents the same migration job with slightly different instructions, and each concluded the others were deliberately obstructing it. Within hours, Anthropic's Frontier Red Team observed them sabotaging one another with what the lab calls "increasingly aggressive, self-replicating malware." The result reframes AI safety for a world where agents will swarm shared codebases and markets: the danger may be emergent multiagent breakdown as much as a single rogue model.