Explore our new Cybersecurity Terms Glossary, explained by our experts.
Cybersecurity Terms Glossary
The OWASP Top 10 for LLM Applications is a community-maintained list of the most critical security risks in applications built on large language models.

What Is the OWASP LLM Top 10?
The OWASP Top 10 for LLM Applications does for AI software what the original OWASP Top 10 did for web applications: it gives developers, testers, and procurement teams a common list of what goes wrong most often. First published in 2023 and substantially revised for the 2025 edition, it is maintained by practitioners reporting from real deployments rather than derived from theory, which is why the entries read like incident summaries.
| ID | Risk | In one line |
|---|---|---|
| LLM01 | Prompt Injection | Crafted input overrides the application’s instructions |
| LLM02 | Sensitive Information Disclosure | The model reveals data it should not, from context, memory, or training |
| LLM03 | Supply Chain | Compromised models, datasets, adapters, or dependencies |
| LLM04 | Data and Model Poisoning | Corrupted training or fine-tuning data implants a behaviour |
| LLM05 | Improper Output Handling | Model output reaches a browser, shell, or database unvalidated |
| LLM06 | Excessive Agency | Too much permission, functionality, or autonomy |
| LLM07 | System Prompt Leakage | Hidden instructions, and the secrets placed in them, are exposed |
| LLM08 | Vector and Embedding Weaknesses | Retrieval and embedding stores leak or are manipulated |
| LLM09 | Misinformation | Confident false output, compounded by over-reliance |
| LLM10 | Unbounded Consumption | Runaway tokens, cost, and resource use |
Three entries are new or substantially reframed, and each reflects what practitioners actually encountered. System prompt leakage was promoted to its own entry after a year in which hidden instructions turned out to contain credentials, business rules, and internal URLs rather than just personality guidance. Vector and embedding weaknesses arrived because retrieval-augmented generation became the default architecture, bringing permission bypass, poisoned documents, and the inversion risks described in our note on embedding inversion. Unbounded consumption replaced the narrower denial-of-service framing to include the commercial reality that token spend is now an attack surface. Meanwhile denial of service and overreliance from the 2023 edition were folded into other entries.
The list earns its value as a working tool in four places. In design review, walk a proposed architecture through each entry and note where a control exists and where it does not—LLM05 and LLM06 usually expose the most. In test scoping, require that an assessment covers all ten explicitly, because a report that only lists prompt injection findings has probably not looked at output handling or consumption limits. In procurement, ask a vendor to state how their product addresses each entry; the quality of the answer is informative regardless of the content. And in developer training, it provides a vocabulary that security and engineering can share, which is most of what makes a secure development process work.
It is a list of common risks, not a standard, a certification, or a complete threat model. It says nothing about model bias or safety, barely touches governance, and does not replace conventional application security testing of the same system—most LLM applications still fail on authentication, authorisation, and dependency management before anything AI-specific. Pair it with MITRE ATLAS for adversary techniques, NIST AI 100-2 for attack taxonomy, and the NIST AI Risk Management Framework for governance. OWASP has also published companion material on agentic systems and on securing generative AI deployments, which extends the coverage towards AI agent security.
Referencing the list is a straightforward way to evidence that you considered known AI risks, which supports UK GDPR accountability and sits comfortably alongside NCSC and CISA guidance on secure AI development. For organisations facing the EU AI Act, it offers a practical reading of what Article 15’s cybersecurity requirement means in engineering terms. At Privia Security, we structure our AI Penetration Test Services around all ten entries plus the conventional application layer beneath them.
The original covers web application vulnerabilities such as broken access control and injection, and it still applies to any AI application, because those are web applications too. The LLM list covers risks introduced by the model layer: instruction manipulation, output that cannot be trusted, excessive agency, and retrieval weaknesses. Use both, not one instead of the other.
Usually LLM06, excessive agency, and LLM05, improper output handling. Both are architectural, both are entirely within your control, and both determine how much damage every other entry can cause. Prompt injection gets more attention, but an injected prompt against a tightly scoped agent with validated output is an inconvenience rather than an incident.
No. It is an awareness and prioritisation resource maintained by a community project, not a standard with a conformity scheme. You can require coverage of it in a penetration test scope or a supplier questionnaire, and reference it in your risk documentation, but nobody certifies compliance with it.
It has been revised roughly annually since 2023, with the 2025 edition introducing system prompt leakage and vector and embedding weaknesses and reworking several others. Given how quickly AI deployment patterns change, check the current version before using it to scope work rather than relying on a copy saved last year.
You May Be Interested In These
We use a few cookies that are needed to make this website work, and they are always on. We would also like to set analytics cookies, to see which pages are useful; functional cookies, so that our live chat works; and advertising cookies, to measure our campaigns. We will not set any of those until you say yes, and you can change your mind at any time from the Cookie Settings link in the footer.
Choose the cookies you are happy for us to use. Strictly necessary cookies are always on; everything else is off until you turn it on.
These cookies keep the website working: they hold your security checks, your form submissions and the choice you make here. The site cannot run without them, so they are always on.
These let our live chat open and remember your conversation, and let the newsletter sign-up work. Switch them off and the chat window will not appear.
These tell us which pages are read and how people find us, so that we can improve the site. The reports are about groups of visitors, never about you personally.
These let us measure our advertising and show you our adverts on other websites, such as LinkedIn, Google and Meta. Switch them off and you will still see adverts, but they will be less relevant.