Last checked: 7 November 2026
In short: AI security is the protection of AI systems, and the data they use, from attack, theft and tampering. It covers threats such as prompt injection, poisoned training data and stolen models.
Table of Contents
How it works
- Map the weak points. Teams list where an AI system could be attacked, from its training data to its prompts and connected tools.
- Build in defences. They limit access, check inputs, protect data and restrict what the AI can do without approval.
- Monitor and respond. Security staff watch for odd behaviour, log activity and patch weak points when attacks appear.
Where you’ll meet it
Security settings in Microsoft Copilot, ChatGPT Enterprise and Gemini control access and data. Security firms now sell tools that scan AI prompts and replies for attacks.
What it means for your business
- AI adds new ways in. A chatbot linked to your files or email can be tricked into leaking them.
- Limit what AI can touch. Give AI tools access only to the data and actions they need.
- Add AI to your security checks. Include AI tools in risk reviews, supplier checks and staff training.
Watch out for
- Prompt injection. Hidden instructions in a web page or email can hijack an AI assistant.
- Unapproved tools. Staff pasting data into free AI apps create risks your IT team cannot see.
Example
National Cyber Security Centre. On 27 November 2023 the UK’s NCSC and the US Cybersecurity and Infrastructure Security Agency published Guidelines for Secure AI System Development. Agencies from 18 countries, including the UK and US, endorsed them. The guidance covers four stages of an AI system’s life: secure design, secure development, secure deployment, and secure operation and maintenance.
AI security vs AI safety
| AI security | AI safety | |
|---|---|---|
| Protects | The AI system and its data | People and society from AI harm |
| Threat comes from | Outside attackers or insiders | The AI’s own behaviour or misuse |
| Typical fix | Access controls and monitoring | Testing, training and guardrails |
Frequently asked questions
Do normal cyber security rules cover AI tools?
Partly. Passwords, access controls and patching still apply. AI adds new risks, such as prompt injection and leaked training data, that need extra checks.
Which AI security risk should a small business tackle first?
Data leaks through staff using unapproved AI tools. Provide an approved business tool, set clear rules on what data may go in and train staff.
Can attackers use AI against my business?
Yes. Criminals use AI to write convincing phishing emails, clone voices and find weak points faster. Train staff to confirm unusual payment requests by phone.