What Is OWASP Top 10 for LLM Applications?
The OWASP Top 10 for LLM Applications is a list of the most critical security risks in software built on large language models, published by the OWASP GenAI Security Project. The current 2025 edition covers risks from prompt injection to unbounded consumption.
Also known as:
The 2025 List
- LLM01 Prompt Injection. Inputs that change the model's behavior in ways the developer did not intend.
- LLM02 Sensitive Information Disclosure. The model or app exposing private data, credentials or proprietary information.
- LLM03 Supply Chain. Risks in third-party models, datasets, plugins and libraries.
- LLM04 Data and Model Poisoning. Tampered training or fine-tuning data that plants backdoors or bias.
- LLM05 Improper Output Handling. Passing model output to other systems without validation, enabling injection downstream.
- LLM06 Excessive Agency. Giving an LLM-based agent more tools, permissions or autonomy than it needs.
- LLM07 System Prompt Leakage. Secrets or security rules placed in system prompts that can be extracted.
- LLM08 Vector and Embedding Weaknesses. Weaknesses in retrieval systems, such as poisoned or leaky vector stores.
- LLM09 Misinformation. Confident but false output that people or systems act on.
- LLM10 Unbounded Consumption. Uncontrolled usage that drives up cost or denies service.
How It Differs From the OWASP Top 10
The classic OWASP Top 10 covers web application risks such as broken access control and injection. The LLM list covers risks that come from adding a model to an application: the model's inputs, outputs, data and the actions it can take. Most LLM apps need both, because the surrounding code still has ordinary web vulnerabilities.
Why It Matters for Agents
As LLM apps become agents that call tools and change systems, risks such as prompt injection, improper output handling and excessive agency stop being about bad answers and start being about bad actions. The OWASP GenAI Security Project publishes separate guidance on agentic AI threats for this reason.
Using the List in Practice
Teams use the list as a checklist for threat modeling and design reviews: limit agent permissions, treat model output as untrusted input, keep secrets out of prompts, validate retrieval sources and set usage limits. It is a starting point for risk discussions, not a complete security standard.
How Nullify Helps
Nullify does not claim to detect every LLM-specific risk. It finds the flaws in the code around the model, such as injection, authentication and authorization failures, IDOR and SSRF, which is where improper output handling and excessive agency usually turn into real exploits. See how Nullify secures the code around your AI features.
Key Takeaways
- The OWASP Top 10 for LLM Applications lists the most critical risks in LLM-based software.
- The 2025 edition runs from LLM01 Prompt Injection to LLM10 Unbounded Consumption.
- It complements the classic OWASP Top 10 rather than replacing it.
- Agent risks such as excessive agency turn bad output into bad actions.
Read More About OWASP Top 10 for LLM Applications
Code Security
See Nullify in Action
Nullify finds the injection and authorization flaws that turn LLM risks into real exploits.