Study finds AI agents will share secrets if they get offered help from others

Cheap coding agents surrendered credentials or ran untrusted scripts when simulated peers made assistance conditional on violating explicit security instructions.