Security for AI systems

Overview

AI adds attack routes such as prompt injection, poisoned data and model theft on top of ordinary security needs. Treat inputs and outputs as untrusted, limit what systems can reach, and test adversarially before launch.

Key risks

• Prompt injection
• Data leakage through outputs
• Poisoned training or retrieval data
• Insecure plugins and integrations

What to do now

• Use the OWASP Top 10 for LLM Applications and MITRE ATLAS as starting references.
• Test adversarially before launch.
• Limit what the system can reach.
• Monitor for abuse.

Questions to ask

• Have we threat-modelled this system and its connections?
• What can the system reach that it does not need?
• Do we validate outputs before they trigger actions?
• Have we tried prompt injection and data extraction?
• Is AI included in our incident plan?

Related tools

• LLM security checklist
• Risk register

Region rules: EU · US · UK · India · Global standards.

Reviewed 3 Oct 2026. Sources: European Commission: AI Act overview.