The AI threat landscape

Lesson 1 of 8 · 30 min · part of the AI Security Professional certification

AI systems inherit every classic security problem — broken auth, leaked secrets, injection — and add new ones, because their behaviour is driven by data and natural language rather than only by code.

The attack surface spans the whole lifecycle:

• Training data — can be poisoned, or can contain sensitive data the model later leaks.
• The model — can be stolen (extraction), probed (inversion, membership inference) or backdoored.
• Prompts and context — can be manipulated by users (direct injection) or by content the model reads (indirect injection).
• Tools and agents — give the model real-world permissions, turning a text trick into an action.
• Outputs — can carry XSS, SQL or shell payloads into downstream systems.

Interactive exercise: Match the incident to the OWASP risk

Interactive exercise: When does the attack happen?