Grant Larsen

Two separate dials control agent risk: how autonomous it is and how sensitive the data it touches.

OAuth 2.0 scopes are the only real defense against cascading agent failures.

Policy-based access control is the only framework built for machines that don't have jobs.

AI agents can inherit permissions that compose into actions none were authorized to take.

Organizations need playbooks built for AI agent failures, not traditional security incidents.

Most enterprises deploying AI agents lack the governance to control them.

Four span types reveal what agents actually hide during execution.

Establish per-agent baselines to distinguish intended change from genuine anomalies.

Agents need per-agent baselines because their behavior has no fixed normal.

Autonomous agents operating inside approved apps pose risks traditional security tools can't detect.

Organizations deploying AI agents lack the logging infrastructure to investigate what they do.

Classical monitoring misses the reasoning loops and behavioral patterns that define agent risk.

Monitoring must watch what agents do between input and output.