The OWASP Top 10 for LLM Applications is the security community's list of the biggest risks in AI apps. This set reimagines the 2026 list as friendly visual cards, a way to start serious conversations about AI security with students, developers, and non-specialists. For awareness and discussion, not a replacement for the official OWASP guidance.
Two lists, paired — because neither covers the other's ground:
- LLM Top 10: risks when AI is part of an app.
- Agentic Top 10: risks when AI can act — with tools, memory across sessions, and real consequences. See the Agentic Top 10 →
LLM risk
Robot reads or writes something risky.
Agentic risk
Robot uses tools, identity, memory, or permissions to do something risky.
Prompt Injection
When hidden instructions sneak into what the robot reads, hears, sees, or remembers.
View poster →
Sensitive Information Disclosure
When the robot reveals secrets it should keep — in its answer, its reasoning, or its training data.
View poster →
Excessive Agency
When the robot is allowed to do too much, with too much power, on its own.
View poster →
Supply Chain
When a model, tool, or package the robot loaded isn't the safe thing you thought.
View poster →
Data and Model Poisoning
When bad training or fine-tuning data teaches the robot the wrong things — and it lasts.
View poster →
Unbounded Consumption
When the robot uses way too much — costing money and grinding things to a halt.
View poster →
Misinformation
When a fluent, confident answer is wrong — and someone acts on it.
View poster →
Hidden Context Exposure
When the robot's backstage instructions, notes, or tool rules slip out. (Renamed from System Prompt Leakage.)
View poster →
Vector and Embedding Weaknesses
When the robot's memory shelves mix things up, leak, or get tricked.
View poster →
Improper Output Handling
When an app trusts the robot's output without checking it first.
View poster →About this set & license
Robot Explains is an independent educational adaptation and is not official OWASP guidance. It is based on the OWASP Top 10 for LLM Applications (2026) by the OWASP GenAI Security Project.
The OWASP material is licensed under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0), and these adapted posters are shared under the same license.