GitHub
AI Governance Toolkit
The public maturity self-assessment and framework from CarbeneAI.
Services
Exercise the AI-enabled SOC. Close the loop.
A closed-loop practice for the AI-enabled SOC.
Audience
Who we serve
Security leaders, SOC teams, executives, and boards at organizations running an AI-enabled SOC, adopting agentic security tools, or putting AI agents into production.
The work
Problems we solve
- 01
Vendor use cases, untested
An AI SOC or agentic security tool shipped with the vendor's use cases, and nobody has tested them against your own risks.
- 02
Agents in production, no plan
AI agents are in production, and the incident response plan does not say who notices, who can shut one off, or who tells the business.
- 03
Findings that stay in the report
Tabletop findings stay in a report. They never become detections, and nobody re-tests them.
- 04
The board has not rehearsed
Executives and the board have not rehearsed AI agent misuse, shadow AI, or an incident at an AI vendor.
In the room
Questions to ask
- 01
When one of our SOC's AI agents fails or is manipulated, who notices?
- 02
Which of our AI-related detections actually fire? How do we know?
- 03
What are our autonomy limits and kill-switch rules for AI in the SOC?
- 04
What does the board do in the first hour of an AI vendor incident?
- 05
When did we last re-test what the last exercise found?
The method
One closed loop
- 01
Exercise
Run the SOC and the business through the incident, including failure or manipulation of the SOC's own AI agents.
- 02
Derive
Write the AI security use cases the exercise exposes.
- 03
Validate
Test them against your telemetry. Keep the ones that fire. Record the ones that do not.
- 04
Decide
Turn results into AI governance controls and AI SOC design decisions, including autonomy limits and kill-switch rules.
- 05
Re-test
Run it again and check that controls and detections hold.
Mapped to
- MITRE ATT&CKAdversary behavior
- MITRE D3FENDDefensive countermeasures
- MITRE ATLASAttacks on AI and ML systems
Public MITRE knowledge bases. ATT&CK for adversary behavior, D3FEND for defensive countermeasures, ATLAS for attacks on AI and ML systems.
Services
Engagements
- 01Flagship · 4 to 6 weeks
Agentic SOC Readiness Sprint
Exercise the SOC, including failure or manipulation of its own AI agents. Derive and validate the use cases in your environment, set autonomy and kill-switch rules, and map evidence to NIST AI RMF and ISO 42001.
- 02Half day
AI Incident Readiness Exercise
Executive and board exercise on AI agent misuse, shadow AI, and AI vendor incidents. After-action report and governance gap list.
- 03Independent review
AI Security Use-Case Audit
Review of existing AI-related detections and agent telemetry. Coverage map, broken or missing use cases, prioritized backlog.
- 04Quarterly
Annual Readiness Program
The loop on a yearly cadence: quarterly exercises, backlog upkeep, re-tests, and board reporting.
The practice
Our team
Our team has delivered tabletops and war games at Booz Allen Hamilton, EY and AT&T DFIR and built a SOC at Accenture.
We run the exercise, and we write and audit the detections.
Cyber Defense Tactics is a practice of CarbeneAI.
Free
Open material.
Three things sit outside the engagements: a tabletop kit, the public library, and the Brief.
- 01 · Free
AI Agent Tabletop Kit
One scenario for an AI-agent incident, tied to NIST AI RMF risks and MITRE ATT&CK techniques. Confirm an address and the file arrives by email. It is separate from the engagements.
- 02 · Public
Reference library
Public pages for adversary behavior, defensive countermeasures, and attacks on AI and ML systems. No account.
- 03 · Monthly
The Cyber Defense & AI Brief
A monthly note. The public record of the method, the reference library, and what exercises show about AI in the SOC.
- 14 tactics
ATT&CK hub
- 7 tactics
D3FEND hub
- ATT&CK ↔ D3FEND
Purple matrix
- L1 to L3
Use case development
- 6 modules
Detection course
- 20 playbooks
IR playbooks
Not ready for a scoping call? Subscribe to the Brief.