The AI Security Lifecycle Checklist
Checklist
A stage-by-stage security check for AI systems, organised around the five lifecycle phases used in the UK AI cyber security code and ETSI EN 304 223.
Secure design
- Threat model the AI component, including hostile inputs and indirect prompt injection.
- Limit what data and tools the system can reach to what the task needs.
- Decide what happens when the model is unavailable or returns unusable output.
Secure development
- Record the source and version of every model, dataset and third-party component.
- Check training and retrieval data for poisoning risk and provenance gaps.
- Keep secrets and credentials out of prompts and system instructions.
Secure deployment
- Treat prompts, system instructions and retrieval sources as configuration under change control.
- Rate-limit and monitor for abuse patterns.
- Confirm users can report harmful or incorrect output.
Maintenance and end of life
- Name who is notified when the model provider changes the underlying model.
- Define what is retested after any model, prompt or data change.
- Plan how logs, embeddings, fine-tuning data and credentials are removed when the system is retired.
Start a conversation
Have a product or AI decision to make?
Useful first calls usually start with one unclear decision, a deadline and a team that needs a practical next move.
Tell us about it