Kim Maida uses an incident-response demonstration to show how broad API keys let an agent take dangerous actions, including deleting a database without checking backups. Asking a person to approve every action does not solve the underlying access problem and can create consent fatigue.
She traces requests through an agent runtime, MCP client, tool server and downstream resource, then explains OAuth token exchange. A security token service checks workload and user identity, evaluates delegated permissions, and issues short-lived credentials limited to the required audience and scopes.
In the policy-controlled demonstration, a certificate task receives narrow access, a destructive database operation is blocked before credentials are minted, and a production restart remains unauthorized even after user consent. The discussion distinguishes existing resource scopes from optional MCP-specific scopes and emphasizes auditable, task-specific attribution.
Watch on YouTube
