UK AI Cyber Security Code of Practice checklist
The UK Department for Science, Innovation and Technology published its Code of Practice for the Cyber Security of AI on 31 January 2025. It is voluntary and sets 13 principles across the life of an AI system, from design to disposal. Here is each principle as a check you can answer, with the evidence to keep.
Who it is written for
The code names the parties responsible for each provision: developers who build or adapt models, system operators who run AI in their services, and data custodians who control the data. One company can hold all three roles for a single system, for example when it fine-tunes a model and runs it in its own product. The code also names end-users and affected entities, but sets no provisions for them to meet. Note which role you hold for each system before you start.
Where to start if you have a month
- Principle 5List your models, data, prompts and configuration, with who can reach each.
- Principle 3Write a threat model for the AI system that matters most.
- Principle 6Check access to model APIs and pipelines, and publish a vulnerability disclosure policy.
- Principle 12Turn on logging of inputs, outputs and actions, and decide who reviews it.
- Principle 1Run a short session on AI-specific attacks for the people who build and run these systems.
Secure design
- 1
Raise awareness of AI security threats and risks
Have the people who build, run and buy your AI systems been taught the AI-specific attacks, such as prompt injection, data poisoning and model inversion?
Evidence: Training material and attendance for developers, operators and system owners.
- 2
Design your AI system for security as well as functionality and performance
Before building, did you write down the business need and the AI security risks, and plan how to reduce them?
Evidence: Security requirements in the design document, signed off before build.
- 3
Evaluate the threats and manage the risks to your AI system
Is there a threat model for each AI system, kept up to date when the model, data or tools change?
Evidence: A threat model with AI-specific threats, a date and an owner, and the decisions taken on each risk.
- 4
Enable human responsibility for AI systems
Can a person oversee what the system does, and do users know what it must not be used for?
Evidence: Human oversight points in the design, and published limits on use.
Secure development
- 5
Identify, track and protect your assets
Do you have a list of your models, data, prompts, weights and their dependencies, with who can reach each?
Evidence: An asset inventory covering models, training data, prompts and configuration.
- 6
Secure your infrastructure
Are access to APIs, models, data and pipelines controlled, and is development kept apart from production?
Evidence: Access controls, environment separation, a published vulnerability disclosure policy and an incident plan.
- 7
Secure your supply chain
Do you assess third-party models, datasets and libraries before use, and re-test when a supplier releases a new model?
Evidence: Supplier assessments, a reason recorded for any poorly documented component, and re-run evaluations.
- 8
Document your data, models and prompts
Is there an audit trail of how the system was designed and changed, including where training data came from and changes to prompts?
Evidence: Data source records, model version history and a change log for prompts and configuration.
- 9
Conduct appropriate testing and evaluation
Is every model and system security-tested before release, including by testers outside the build team?
Evidence: Test reports before each release, and results from independent testing.
Secure deployment
- 10
Communication and processes associated with End-users and Affected Entities
Do users know how their data is used and where the system's limits are, and will you tell them about security updates?
Evidence: User-facing documentation on data use, limits and how updates are communicated.
Secure maintenance
- 11
Maintain regular security updates, patches and mitigations
Is there a way to patch and update the system, and are major model updates tested again before release?
Evidence: Update and patch records, and evaluation results for each major model update.
- 12
Monitor your system's behaviour
Do you log inputs, outputs and actions, and watch for drift, misuse and signs of attack?
Evidence: Logging in place, alerts defined, and a record of what was reviewed.
Secure end of life
- 13
Ensure proper data and model disposal
When a model or dataset is retired or handed over, is it disposed of securely with the data owner involved?
Evidence: A retirement checklist and records of disposal for retired models and datasets.
Also working to NIST AI RMF?
Much of the evidence for the UK code also answers NIST AI RMF 1.0 categories, so one set of records can serve both. This table is our reading. We found no mapping between the two from DSIT or NIST.
| UK principle | NIST AI RMF categories | Shared evidence |
|---|---|---|
| 1 Raise awareness of AI security threats and risks | GOVERN 2, GOVERN 4 | Training records and threat-update bulletins |
| 2 Design your AI system for security as well as functionality and performance | GOVERN 1, MAP 1 | Written requirements and the system's purpose |
| 3 Evaluate the threats and manage the risks to your AI system | MAP 4, MAP 5, MEASURE 3, MANAGE 1 | The threat model, impact assessment, risk register and accepted residual risks |
| 4 Enable human responsibility for AI systems | GOVERN 3, MAP 2, MAP 3 | Oversight points, output explanations and published prohibited uses |
| 5 Identify, track and protect your assets | GOVERN 1, MAP 4, MANAGE 4 | The AI inventory, component list and recovery plans |
| 6 Secure your infrastructure | GOVERN 4, MEASURE 2, MANAGE 4 | Access reviews, vulnerability disclosure policy, incident and recovery plans |
| 7 Secure your supply chain | GOVERN 6, MAP 4, MANAGE 3 | Supplier reviews and third-party risks |
| 8 Document your data, models and prompts | MAP 2, MAP 4, MEASURE 2 | Data sources, model versions, known limits and change logs |
| 9 Conduct appropriate testing and evaluation | MEASURE 1, MEASURE 2, MEASURE 4 | Test plans, results and reviews of the metrics |
| 10 Communication and processes associated with End-users and Affected Entities | MAP 2, MEASURE 2, MANAGE 4 | User guidance on data use and limits, update notices, incident support |
| 11 Maintain regular security updates, patches and mitigations | MEASURE 2, MANAGE 3, MANAGE 4 | Update records and re-tests after major model updates |
| 12 Monitor your system's behaviour | MEASURE 2, MEASURE 3, MANAGE 4 | Monitoring, alerts and incident records |
| 13 Ensure proper data and model disposal | GOVERN 1, MANAGE 4 | Retirement and disposal records |
Want to know where you stand?
Our free assessment includes the UK AI Cyber Security Code of Practice. Answer the questions and you get a short report with the gaps that matter most and what to do first. The same assessment covers NIST AI RMF 1.0, the OWASP Top 10 for LLM Applications and MITRE ATLAS.
Source
Checked against the source on 15 November 2026.
Principle titles are as published. The checks and evidence are our plain-English reading of each principle's provisions. Free to use and adapt inside your organisation. Not legal advice.