Building with AI introduces new security risks. This reuses the core from Cyber Security Fundamentals: AI Security, applied to building AI systems.
AI Security
Securing AI — and defending against AI-powered attacks.
Tap or hover a part to learn more.
OWASP LLM Top 10.
AI systems face new risks: prompt injection (manipulating inputs), data poisoning (corrupting training data), model theft and sensitive-data leakage. OWASP's LLM Top 10 is the reference.
Check your understanding
1. What is prompt injection?
2. What reference lists AI/LLM risks?
Keep learning
Key risks when building AI
- Prompt injection — malicious input that hijacks an AI's instructions (especially dangerous for agents with tool access).
- Data leakage — sensitive data pasted into prompts or returned in outputs.
- Data poisoning & model theft — corrupting training data or stealing the model.
- The OWASP Top 10 for LLMs — the reference for securing LLM applications.
Controls
Validate and constrain inputs and outputs, apply least privilege to any tools an agent can call, never trust model output blindly, and keep sensitive data out of untrusted models. Security must be designed into AI systems from the start — the same secure-by-design principle.
