272 controls across 14 families, with practical guidance for governing, securing and maintaining human control over AI.
The AI North Star Controls Catalogue is now available here on the site. It brings together the requirements, implementation guidance and verification criteria organizations need to assess their AI systems and strengthen the safeguards around them.
AI control involves decisions about what systems can access, which actions they can take, who approves those actions and how people intervene when something goes wrong. The catalogue turns those questions into specific controls with assigned responsibilities, implementation steps and evidence of effectiveness.
Developed alongside The Unconstrained Machine, it is a standalone resource for security professionals, governance teams, developers, auditors and business leaders responsible for AI.
How the catalogue was developed
The catalogue was assembled by reviewing and consolidating controls and requirements from established AI governance, security and regulatory sources, adapting cloud-oriented material for broader deployment environments and adding safeguards addressing human authority and autonomous systems. CSA’s AI Controls Matrix provided a substantial control foundation. NIST’s AI Risk Management Framework informed the organization of risk-management activities across the lifecycle, while ISO/IEC 42001 contributed management-system requirements for accountability, assessment and continual improvement. The EU AI Act and GDPR supplied references for applicable legal obligations, including oversight, transparency, data protection and individual rights. MITRE ATLAS and the OWASP LLM Top 10 informed adversarial threats and application-security safeguards. These sources serve different purposes; the catalogue brings their relevant requirements and practices into a consistent structure, with mappings retained on individual control pages.
From requirements to implementation
The catalogue contains 272 controls: 191 core and 81 supplemental. Every control has its own page explaining the requirement, why it matters, when it applies and who should own it.
Each page includes:
- Implementation steps.
- Verification tests and expected results.
- Evidence to retain.
- Actions to take when the control fails.
- A hypothetical example illustrating its use.
- Related controls and framework mappings.
The objective is to give practitioners enough detail to move from an expectation such as “maintain human oversight” to an implemented process with assigned authority, functioning intervention mechanisms and evidence that those mechanisms work.
Core and supplemental designations help organize selection and depth. Applicability depends on the system, its operating environment and the consequences of failure.
Four types of control working together
The catalogue distinguishes four ways control is exercised:
- Technical controls: hard limits built into the system, including access restrictions, execution boundaries and enforced resource limits.
- Governance controls: hard external boundaries established through laws, regulations and binding rules.
- Operational controls: policies, procedures, responsibilities and working practices that guide operation, review and response.
- Contextual controls: information and environmental conditions that shape behavior and decisions.
An agent authorized to issue refunds illustrates how these types work together. Technical controls enforce permissions and transaction limits. Operational controls establish approval and escalation procedures. Contextual controls make the proposed action, supporting information and consequences clear to reviewers. Governance controls establish applicable external obligations.
A control family organizes a subject, such as privacy or human oversight. The four types describe how control operates within that subject.
Coverage across 14 families
The catalogue covers the AI lifecycle through 14 control families:
- Lifecycle Governance
- Fairness, Bias & Rights Impact
- AI Supply Chain & Third-Party Risk
- Auditability & Evidence
- Data Governance & Privacy
- Transparency & Explainability
- Human Oversight & Escalation
- Testing & Robustness
- Monitoring & Drift Detection
- Recourse & Incident Response
- Objective & Policy Constraints
- Security & Abuse Resistance
- Model & Adversarial Security
- Agentic & Autonomous System Controls
This coverage connects organizational responsibilities with the technical and operational safeguards needed in deployed systems. It includes data handling and retrieval permissions, model evaluation and integrity, meaningful review, incident response and recourse for affected people.
The agentic controls extend into delegation, task boundaries, resource acquisition, replication, persistent memory, action reversibility and termination. These become essential considerations when AI systems can use tools and act across connected services. Stopping an agent, for example, requires accounting for delegated workers, queued actions and external requests already in progress.
How to start using the catalogue
Begin with one AI system your organization already uses. Identify what it can access, what it can change, who could be affected and who has authority over its operation. Then use the relevant families to examine the safeguards around it.
For each applicable control, assign an owner, inspect the implementation and test the expected result. Record gaps and prioritize improvements according to their consequences. An agent with payment authority warrants different safeguards from an assistant summarizing public documents.
Use the control identifiers when documenting architecture decisions, assessment findings, implementation tasks and evidence. The consistent structure gives business, governance and engineering teams a shared basis for discussing what is required and whether it works.
The complete catalogue is freely accessible. Browse the 14 control families and explore the individual controls.
