Securing Autonomous AI Agents
Opening keynote · SF AppSec & DevSecOps Summit · May 14, 2026 · San Francisco
You cannot prevent an agent from being manipulated, but you can make it not matter when it is. The control plane is the blast radius.
The talk lays out five blast-radius controls for autonomous agents. Each one has its own page with a code example:
- 01Capability scopingLeast privilege per invocation, not per agent.
- 02Sandboxed executionDefault-deny everything outside the sandbox.
- 03Human-in-the-loop gatingCurate irreversible actions. Gate them in code, not policy.
- 04Behavioral auditTrust syscalls, not narratives.
- 05Agent identityFirst-class principal. Not a shared service account.