AI Security for Cyber Securitys
AI security addresses threats to systems built on machine learning. For large language model applications, risks include prompt injection, data leakage and excessive agent permissions. For ML pipelines and models, risks include poisoned training data, model theft and adversarial inputs, so data, code and artefacts need controls.
Stage: Cloud, Application and AI Security · Part of the Cyber Security roadmap · Last reviewed 19 November 2026
What to learn in AI Security
- Demonstrate a prompt injection against a simple chatbot
- Limit an agent's tool permissions to what its task needs
- Describe training data poisoning and how provenance checks reduce it
- Protect model files and API keys in an ML pipeline
- Validate and filter model output before it reaches other systems
AI Security interview questions
What is prompt injection?
Prompt injection is when attacker-controlled text, from a user or from content the model reads, contains instructions that override the developer's intent. Because models do not reliably separate instructions from data, apps need least-privilege tools and output checks.
Why are AI agents with tool access a security risk?
An agent can take real actions, such as sending email or running code, so a successful injection can cause harm beyond text. Limit tool permissions, require human approval for sensitive actions and log what the agent does.
Where AI Security fits in the Cyber Security roadmap
AI Security belongs to the Cloud, Application and AI Security stage (beginner level), which takes about 5 hours. Where the risk moved.
Other units in this stage
Back to the full Cyber Security roadmap