UK publishes the AI Cyber Security Code of Practice

The UK government published its AI Cyber Security Code of Practice on 31 January 2025, a voluntary set of baseline security principles for AI systems and the organisations that build and run them. It was released with an implementation guide and is meant to seed a global standard at the European Telecommunications Standards Institute.

What it is

The code treats AI as different from ordinary software because of risks such as data poisoning, model obfuscation and indirect prompt injection. It builds on the NCSC guidelines for secure AI development published in November 2023 and is framed as an addendum to the UK Software Code of Practice. In a call for views, 80% of respondents backed the approach. The government planned to submit the code as the basis for ETSI standard TS 104 223 with guide TR 104 128.

What it requires

Its 13 principles run across five lifecycle phases: secure design, development, deployment, maintenance and end of life. Provisions use shall for requirements and should for recommendations. Examples: threat modelling must address AI-specific attacks such as data poisoning, model inversion and membership inference; developers shall release cryptographic hashes so others can verify model components; developers and system operators shall publish a vulnerability disclosure policy; and system operators shall log system and user actions for investigations.

What security leaders should do

Even outside the UK, the code offers a concrete, public control set for AI security that can sit beside an existing secure development programme. Work out whether your organisation is a developer, a system operator, a data custodian or several of these, then run a gap assessment against the matching principles. Start with the items that reuse existing processes: asset inventory, vulnerability disclosure, logging and incident plans that name AI systems explicitly.