What it is about
This list comes from the OWASP GenAI Security Project, a community of practitioners who build and test AI apps. It ranks the risks that show up again and again in real systems: prompt injection, data leaks, poisoned training and retrieval data, unsafe model output, too much freedom for the model, supply chain problems, and runaway cost. Each risk comes with ideas for how to test for it and reduce it.
Why it matters
Most LLM apps fail at the same few places, and attackers know the list as well as defenders do. If your team can explain these ten risks and show how each one is handled, you are already ahead of most AI projects. Start with prompt injection, output handling, and limiting what the model is allowed to do.
Who should read it
Developers building chatbots and AI agents, penetration testers, AI red teamers, and security reviewers.
Source
Original document by OWASP GenAI Security Project.
The original document is not hosted here. Our note above is our own summary and opinion, in simple words.
Want to learn this properly?
Go deeper with Cyber Legends training and services.
