AI security guide · Updated
The OWASP Top 10 for LLM applications, explained
The OWASP Top 10 for LLM Applications is the most widely used list of security risks for software built on large language models. The 2025 edition, published November 17, 2024, runs from prompt injection (LLM01) to unbounded consumption (LLM10). Below is each risk in plain English, with what it means for a real app and the first thing to do about it.
By George Shvaya, founder of Webvello · Published
Key takeaways
- The OWASP Top 10 for LLM Applications 2025 was published on November 17, 2024, and prompt injection is LLM01.
- Three of the ten risks are about what the model is allowed to do with its output and tools: Improper Output Handling, Excessive Agency and Unbounded Consumption.
- System Prompt Leakage (LLM07) and Vector and Embedding Weaknesses (LLM08) were new in the 2025 edition.
- OWASP published a separate Top 10 for Agentic Applications for 2026 on December 9, 2025, for systems where AI agents act on their own.
- Most of the fixes are ordinary security engineering: least privilege, input and output validation, access control and rate limits.
What are the OWASP Top 10 risks for LLM applications?
| ID | Risk | What it means | First defense |
|---|---|---|---|
| LLM01:2025 | Prompt Injection | Input or content the model reads changes its behavior in unintended ways. | Limit what the model can do; require approval for consequential actions. |
| LLM02:2025 | Sensitive Information Disclosure | The model reveals personal data, credentials or confidential business data. | Keep secrets out of context; filter what goes in and comes out. |
| LLM03:2025 | Supply Chain | Third-party models, datasets, plugins or packages are compromised or vulnerable. | Vet and pin models and dependencies; track their advisories. |
| LLM04:2025 | Data and Model Poisoning | Training, fine-tuning or retrieval data is manipulated to plant behavior. | Control and check data sources; test for unexpected behavior. |
| LLM05:2025 | Improper Output Handling | Model output reaches a browser, shell, database or API without validation. | Treat output as untrusted input: encode, validate, parameterize. |
| LLM06:2025 | Excessive Agency | An agent has more tools, permissions or autonomy than its task needs. | Least privilege per tool; human approval for high-impact steps. |
| LLM07:2025 | System Prompt Leakage | Instructions or secrets in the system prompt are exposed. | Never put credentials or security controls in the system prompt. |
| LLM08:2025 | Vector and Embedding Weaknesses | Retrieval systems leak data across users or accept poisoned content. | Enforce access control in retrieval; validate what gets indexed. |
| LLM09:2025 | Misinformation | The model produces confident, false or unsupported output that people rely on. | Ground answers in sources; show citations; keep a person in the loop. |
| LLM10:2025 | Unbounded Consumption | Uncontrolled use drives cost, denial of service or model extraction. | Rate limits, quotas, input size limits and spend alerts. |
Source: OWASP GenAI Security Project, LLM Top 10 and the 2025 edition. OWASP’s own pages go deeper on each entry, with example scenarios and prevention steps.
Which LLM risks should most teams fix first?
Excessive Agency (LLM06)
Prompt Injection (LLM01)
Improper Output Handling (LLM05)
Sensitive Information Disclosure (LLM02)
Want your AI app checked against the list?
Tell us what your app does with a model. We will tell you which of the ten risks apply and where to start.
Is there an OWASP list for AI agents?
| Publication | Published | What it covers |
|---|---|---|
| Agentic AI – Threats and Mitigations | February 17, 2025 | A threat model for AI agents that plan, use tools and keep memory |
| OWASP Top 10 for Agentic Applications for 2026 | December 9, 2025 | A ranked list of risks for agentic systems |
Agents that connect to tools through the Model Context Protocol add protocol-level risks such as token passthrough and tool poisoning; see MCP security risks.
How should a team use the OWASP LLM Top 10?
- Map each feature that uses a model to the risks that apply to it; a summarizer and an agent that sends email have different lists.
- For each applicable risk, write down the control and who owns it.
- Test the controls with realistic content before launch and after major model or prompt changes.
- Recheck when OWASP updates the list; the 2025 edition changed several entries from 2023.
Related services and resources
Pages that go deeper on the pieces this one touches.
Common questions
Straight answers to what owners ask us first.
What is number one on the OWASP LLM Top 10?
Prompt Injection, listed as LLM01:2025. It was also first in the earlier edition.
Is the OWASP LLM Top 10 a standard we can be certified against?
No. It is an awareness and guidance document from the OWASP GenAI Security Project. Teams use it to structure threat modeling and testing, but there is no certification against it.
Does the OWASP LLM Top 10 apply to apps that use ChatGPT or Claude through an API?
Yes. Most entries are about how your application uses a model: what you put in the prompt, what tools you connect and what you do with the output. They apply whether you host the model or call a provider’s API.
Can you review our app against the OWASP LLM Top 10?
Yes. Our AI application security review walks through each applicable risk, tests the main ones, and gives you a prioritized fix list.
Turn the list into fixes
We review AI applications against the OWASP LLM Top 10 and help fix what we find.