AI agents can ship code, send emails, and change production. How do you keep that power under control? Drawing on real incidents, this talk gives startup teams four practical safeguards: Constrain, Pace, Observe, and Recover, to limit damage, verify outcomes, and recover when things go wrong.
Download the slides (PDF). The slides are available under CC BY-SA 4.0; primary sources are collected in the references at the end.
Four safeguards
- Constrain. Give each agent one narrowly scoped task and only the tools, data and permissions it needs. Enforce the boundaries through infrastructure the agent cannot bypass.
- Pace. Put review before consequential actions, bind approval to the exact action, and cap unattended work, retries and spending.
- Observe. Verify actual tool outcomes with independent evidence. Protect the logs and checks, and continuously test the complete application against adversarial inputs.
- Recover. Keep an independent stop control and protected backups. Rehearse restoration and verify the result. Restoring a database cannot recall disclosed information.
These are informed by established security and AI risk-management guidance.
Further reading
For follow-up questions, add me on LinkedIn. You can also browse my other talks.