The OWASP Top 10 for LLM Applications is a community-maintained list of the most critical security risks for applications built on large language models. It is the best starting point for anyone building, testing or buying an AI application. This article explains the 2025 list in plain language.
LLM01: Prompt Injection
An attacker supplies input that changes what the model does. It can be typed directly by a user or hidden in content the model reads, such as a web page or document. It is the best-known LLM risk.
LLM02: Sensitive Information Disclosure
The model reveals personal data, secrets or proprietary information, from its training data, its prompt or the documents it can retrieve.
LLM03: Supply Chain
Models, datasets, plugins and libraries come from third parties. A compromised component can compromise your application.
LLM04: Data and Model Poisoning
Manipulated training, fine-tuning or retrieval data changes model behaviour, introduces bias or plants backdoors.
LLM05: Improper Output Handling
Model output is passed to other systems without validation. It can then trigger injection, script execution or other downstream flaws.
LLM06: Excessive Agency
An agent is given too many tools, permissions or autonomy, so a manipulated model can do real damage.
LLM07: System Prompt Leakage
Hidden instructions are exposed. This matters when they contain secrets or reveal business logic and security assumptions.
LLM08: Vector and Embedding Weaknesses
Retrieval-augmented generation introduces its own attack surface: poisoned documents, weak access control on a vector store and leakage between users.
LLM09: Misinformation
Models can produce confident but false output, and people rely on it. Overreliance turns hallucination into a security and business risk.
LLM10: Unbounded Consumption
Uncontrolled use of the model can cause denial of service, runaway cost or model theft through extraction.
What to Do Next
Map your application against the list, decide which risks apply and test them. Our AI Security course teaches this hands-on, and our LLM pentesting service tests real applications against the list.
