Runtime Control for AI Agents Fails the Moment Identity Disappears
OWASP LLM-08 is called "Excessive Agency." Read carefully: it is not a model safety problem. It is an identity and authorization problem.
Runtime guardrails — prompt filters, content policies, in-process tool allowlists — help developers build agents that try to behave well. They do not prove agents did behave well when money moves or accounts are deleted.
When identity disappears
The failure mode is consistent: an agent acts with a service account, a broad OAuth scope, or a shared runtime identity. The delegation chain from human to agent to tool is lost. Audit asks who authorized the action — and the answer is a log line, not proof.
Infrastructure enforcement requires identity
Governed execution binds each consequential action to bounded work, a named owner, current policy, and verifiable evidence. Kill switches and budget gates only work when authority is constructed at the identity layer — not bolted on after the agent already holds credentials.
Related
Agent Governance product →