AI security: your questions answered

It means protecting the AI systems you build or use, and using AI safely to support defence[3]. The government's code of practice treats AI security as a subset of cyber security[6].

The NCSC points to prompt injection, data poisoning, hallucinations and bias[2]. OWASP's 2025 list ranks prompt injection first and sensitive information disclosure second[8].

Probably not completely. The NCSC says today's models keep no firm line between instructions and data, so the goal is lowering likelihood and impact[1]. Some use cases may not tolerate what is left[1].

Yes. The government published a voluntary AI cyber security code of practice with 13 principles on 31 January 2025[5]. The NCSC has published guidelines for secure AI system development[7].

The 2025/2026 government survey counted 31% of businesses, and a quarter of charities, as using AI, adopting it or considering it[10]. Among businesses using or adopting AI, 24% had cyber security practices for AI-related risks[10].

It is staff using AI tools the organisation has not approved. The NCSC advises understanding why they do it so the risks can be managed[3].

The NCSC encourages using AI to strengthen defences and notes that suppliers already use it to find and fix vulnerabilities[3]. Output should still be checked, because AI models can produce confident but wrong answers[2].

The NCSC advises scepticism towards vendors who claim to stop prompt injection, and says deny-lists can be rephrased around[1]. Favour suppliers who show how they reduce the risk.

Make a list of your AI tools, with the data and actions each one can reach[3]. Then cut access to what is needed, add human approval for high-risk actions and log activity[9].

The NCSC asks organisations to decide who is accountable for AI security and what the worst-case failure would look like[2]. It also says leaders should have a response plan[2].