Agentic Denial of Service (ADoS) describes a risk in which an AI agent reduces system availability or performance by consuming too many resources or triggering excessive actions. Unlike a conventional denial-of-service attack driven by a flood of incoming traffic, ADoS can arise from the agent’s own activity, such as repeated tool or API calls, retries, or workflows that loop or branch excessively. It can consume compute, API quotas, or processing capacity and delay other work.
This matters wherever agents can act on security, IT, or cloud systems. Defenses include setting per-session quotas and execution-time limits, applying rate limits and circuit breakers, restricting permissions, and requiring human approval for high-impact actions. Monitoring and audit logs can help identify abnormal patterns of agent activity. These safeguards contain runaway automation whether it results from a mistake, a faulty workflow, or misuse.