UK NCSC and US CISA publish Guidelines for Secure AI System Development
On 27 November 2023 the UK NCSC and US CISA published joint guidelines for building AI systems securely, endorsed by agencies from 18 countries.
- Date
- 2023-11-27
- Instrument
- Guidelines for Secure AI System Development
- Published
- Sources
- 2
Compiled September 27, 2026 from public records. The events described took place on November 27, 2023.
By the cisonews.ai Desk · Published
The Guidelines for Secure AI System Development are a joint security guide for anyone building systems that use AI, published by the UK National Cyber Security Centre (NCSC) on 27 November 2023. The NCSC developed them with the US Cybersecurity and Infrastructure Security Agency (CISA), and agencies from 18 countries, the US among them, endorsed the text.
What it is
The document is voluntary guidance rather than law. The NCSC described it as the first guidelines of their kind to be agreed globally, written with industry experts and with 21 other international agencies and ministries, including those of every G7 member. Its stated aim is to help providers build AI systems that work as intended, stay available, and do not reveal sensitive data to unauthorised parties.
What it requires
The guidance is organised around four stages of the AI life cycle: secure design, secure development, secure deployment, and secure operation and maintenance. Design covers threat modelling and trade-offs in system and model design. Development covers supply chain security, documentation, and asset and technical debt management. Deployment covers protecting infrastructure and models from compromise and building incident management processes, and operation covers logging, monitoring, update management and information sharing. The text is aimed mainly at providers who use models hosted by an organisation or reached through external APIs.
What security leaders should do
Treat the four stages as a checklist for any internal team shipping an AI feature, including teams that only wrap a vendor model behind an API. Map each stage to an owner and to controls you already run for conventional software, then note the gaps that are specific to models: training data provenance, model files as supply chain artefacts, and monitoring of model inputs and outputs. Because the text is co-sealed by many national agencies, it also works as a shared vocabulary when questioning AI suppliers about how they build and operate their products.
Sources (2)
- [1]UK National Cyber Security Centrencsc.gov.ukNovember 27, 2023tier A · Primary
- [2]UK National Cyber Security Centrencsc.gov.ukNovember 27, 2023tier A · Primary
Spotted an error? Request a correction. Every fact on this page is tied to a listed source.