Agentic AI risk management standards profile
Risk areas: discrimination and toxicity, privacy and security (memory leakage, cascading compromise), cascading misinformation between agents, misuse (automated cyber or biological attack stages, swarm influence campaigns), reduced human oversight and over-reliance, loss of control (oversight subversion, actions outrunning monitoring), socioeconomic harms, and failures such as self-replication, reward hacking, collusion and deceptive behaviour.
L0lowest of the profile's autonomy levels: the user has direct control with no agent support