OWASP's agentic security Top 10 turns 'excessive agency' into a least-agency discipline

OWASP published a Top 10 for agentic applications that turns runaway autonomy into a named, testable failure. It reframes its 'excessive agency' risk — an agent doing more than it should — into a least-agency discipline: autonomy is something an agent earns rather than gets by default. Its concrete countermeasures are approval gates on state-changing actions, hard step and recursion budgets, and monitors that pause an agent at a policy boundary. It sits alongside OWASP's LLM Top 10, which ranks prompt injection as the number-one risk to any app built on a language model.

What it means If your product lets an agent take actions, this is the checklist to threat-model against: give it the least autonomy the job needs, gate the state-changing steps behind a human, and cap how long it can loop.