You have clicked it. We all have. Somewhere, late at night, between the tenth and the twentieth permission prompt, you click "Always allow", and you have removed yourself from the loop.
What remains is an autonomous agent with no identity of its own, acting on a static API key or a shared service account, with your permissions.

The approval prompt was doing the job of an identity system; with the human gone, nothing is left underneath.
Note that, OWASP now ranks Excessive Agency #3 among LLM risks.

This session opens with that fatal click.
Then, relying on open standards such as SPIFFE and OAuth 2.1 token exchange, we will see how to prevent excessive permissions for agentic workloads, in two steps: workload identity to prove who the agent is, and on-behalf-of delegation to prove on whose behalf it acts and what it may do.
Live demos on a fully open source stack: an agent without identity turned away, an agent without a mandate blocked, and an audit trail that finally answers "who did what, on whose behalf".
You will leave with a clear mental map of the agent identity layers, and a practical answer to give the next time someone asks to click "Always allow".
No IAM expertise required.