Agentic AI Control Baseline · Version 2.0
A control baseline for AI agents that act.
A 20 page security guide with a threat model, 32 controls with the evidence to look for, 14 detections, an incident playbook and a tabletop drill. Four control layers you can test this week, each enforced outside the model.
For CISOs, security architects, SOC and IAM teams, risk, compliance and internal audit.
Get the guide by e-mail
Inside the guide
Controls you can test this week, not a policy to file.
Every control comes with the evidence an assessor should see, and every threat is mapped to OWASP and MITRE. Use it for a 60 minute leadership check, a two week control assessment or a 90 day implementation.
- Reference architecture: six trust boundaries and where each control layer sits
- Threat model: 12 agent threats with OWASP and MITRE references
- Agent tiers: autonomy and blast radius decide the controls each agent needs
- 32 controls: with the evidence to look for and a 0 to 4 maturity score
- 14 detections and an incident playbook for agent compromise
- The 2 a.m. drill, 20 provider questions, a framework crosswalk and a 90 day plan
- Companion Excel workbook on request: agent register with tiering, scored baseline and trackers. Ask for it
Next step
Want a figure for your own organisation?
The AI Risk Assessment shows what your use of AI could cost in euros, which controls are missing, and which fixes pay back first.
