A set of recommended security controls for building, deploying, and operating AI systems safely. It typically covers governance, access control, data protection, model integrity, logging, incident response, and supply chain risk. In practice, it translates AI risk into operational requirements that security, engineering, and compliance teams can apply consistently.
What the code of practice is for
An AI cyber security code of practice is a control-oriented reference point, not a product or a law. It turns broad AI risk into practical expectations for how systems should be designed, deployed, monitored, and governed across the AI lifecycle.
Its value is that it gives security, engineering, and compliance teams a common operational language. That usually makes it easier to decide which safeguards are mandatory, which are risk-based, and which need to be verified before a system goes live.
What it usually covers
Most codes of practice group requirements around a few repeatable control areas: governance and accountability, access control, data protection, model and prompt integrity, logging, incident response, and supply-chain assurance. Those topics matter because AI systems can fail through the same basic control weaknesses that affect other software, but often at greater speed and scale.
In practice, this means the code should help answer questions such as who can change the system, what data it may use, how outputs are monitored, and what evidence exists that controls are actually working. A useful code is specific enough to be applied, but flexible enough to fit different AI architectures and deployment patterns.
How it shapes AI security controls
The strongest codes of practice do more than repeat general security advice. They connect AI-specific failure modes, such as prompt manipulation, model tampering, unsafe tool use, training-data contamination, and insecure integrations, to concrete security requirements.
That makes them useful as a bridge between policy and implementation. Teams can map the code to NIST AI Risk Management Framework principles for governance, to NIST Cybersecurity Framework 2.0 for cross-cutting security outcomes, and to NIST Privacy Framework where AI processing creates privacy risk. It can also align with NIST SP 800-53 Rev 5 Security and Privacy Controls when the organisation needs a formal control catalogue.
Because AI systems often expose APIs, external tools, and automated workflows, the code should also reinforce least privilege, secure defaults, and evidence-driven validation rather than trusting model behaviour alone.
Why it matters in real deployments
A code of practice matters because AI risk is rarely confined to the model itself. Exposure often appears in surrounding systems, such as data pipelines, secrets handling, third-party services, logging, and access paths that were added for speed rather than control.
For that reason, practitioners often use a code of practice to separate baseline controls from higher-risk exceptions. It becomes the reference point for deciding when an AI use case needs stronger review, stronger logging, tighter data boundaries, or more explicit approval before production rollout.
Risk and Threat Considerations
AI security codes fail when they stay too generic or are treated as documentation instead of enforceable control intent. The main risk is that teams assume AI-specific safeguards exist because a policy exists, while the actual system still contains weak data handling, excessive access, poor logging, or unsafe third-party dependencies.
Failure mechanism: Attackers and internal users can exploit the gap between written guidance and real implementation, especially where models, tools, secrets, and data stores are loosely connected.
Impact: That gap can lead to prompt injection, data leakage, model tampering, unauthorised actions, and weak incident containment, particularly when the AI system is integrated into high-trust business workflows.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST AI RMF, NIST CSF 2.0 and CIS Controls v8 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST AI RMF | Govern | AI codes of practice operationalise AI risk governance and accountability. |
| Recommendation — Use governance outcomes to turn the code into owned AI risk controls. | ||
| NIST CSF 2.0 | GV.RM-01 — Risk Management Strategy | The code translates AI risk into a repeatable security risk strategy. |
| PR.DS-01 — Data Management | The code covers data handling, protection, and AI training/usage boundaries. | |
| PR.AA-05 — Least Privilege for Access Authorizations and Entitlements | Access control is central to safe AI system operation and tool use. | |
| Recommendation — Align AI control requirements to enterprise risk management decisions. Define data handling rules for AI inputs, outputs, and training sources. Restrict AI system access, entitlements, and tool permissions to least privilege. | ||
| CIS Controls v8 | CIS-6 — Access Control Management | AI codes commonly require disciplined access control across systems and data. |
| Recommendation — Apply access control management to AI platforms, data, and supporting services. | ||
Practitioner Guidance
Governance implication: Treat the code as a minimum security baseline that must be translated into reviewable engineering and operational controls. If a proposed AI system cannot show how it meets the code, the gap should be resolved before deployment, not after the first incident.
What to watch for: The warning sign is a code that reads well but cannot be traced to owners, test evidence, logging coverage, approval criteria, or rollback procedures. In practice, the best codes are the ones teams can actually inspect, test, and enforce.
Related resources from NHI Mgmt Group
- What is the difference between the UK Code of Practice for AI Cyber Security and the EU AI Act?
- How should security teams prioritise cyber threats across code, dependencies, pipelines, and AI tools in the SDLC?
- How should organisations implement AI security governance when using the UK Code of Practice as a baseline rather than a binding law?
- How should security teams govern AI-generated code in production environments?
Deepen Your Knowledge
Reviewed and updated by the NHIMG editorial team on September 24, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org
