The OWASP Top 10 for LLM Applications is a community-maintained ranking of the ten most critical security risks in software that uses large language models, published by the Open Worldwide Application Security Project. In AI safety and application security practice, it serves as a shared vocabulary for threat modeling, design reviews and testing of chatbots, retrieval systems and AI agents.
The list does not describe model alignment in the abstract. It focuses on how LLM features fail inside real applications: untrusted text reaching the model, model output reaching other systems, and models receiving more permissions than they need. It is maintained by a volunteer working group under the OWASP GenAI Security Project and is revised as attack techniques and deployment patterns change.
What the OWASP Top 10 for LLM Applications covers
The list catalogs application-level risks, meaning weaknesses that appear when a language model is connected to users, data stores, tools and downstream code. Each entry has an identifier, a description, common vulnerability examples, prevention and mitigation strategies, and sample attack scenarios, which makes it usable as a checklist during design and audit.
The current edition, released for 2025, reflects lessons from production deployments of retrieval-augmented generation and agentic workflows. The following table summarizes its ten entries.
| ID | Risk | Core concern |
|---|---|---|
| LLM01 | Prompt injection | Crafted input overrides intended instructions |
| LLM02 | Sensitive information disclosure | Model reveals personal, proprietary or secret data |
| LLM03 | Supply chain | Compromised models, datasets, adapters or libraries |
| LLM04 | Data and model poisoning | Manipulated training or fine-tuning data |
| LLM05 | Improper output handling | Unvalidated output passed to browsers, shells or databases |
| LLM06 | Excessive agency | Too many tools, permissions or autonomous actions |
| LLM07 | System prompt leakage | Secrets or logic exposed through the system prompt |
| LLM08 | Vector and embedding weaknesses | Flaws in retrieval stores and embedding pipelines |
| LLM09 | Misinformation | Confident but false output relied upon by users |
| LLM10 | Unbounded consumption | Uncontrolled cost, resource use or model extraction |
How the list evolved between editions
The first version appeared in 2023, shortly after LLM-based chat interfaces became mainstream, and a minor revision followed the same year. The 2025 edition reorganized several entries because practitioners reported that some original categories overlapped, while new attack surfaces such as vector databases and system prompts had become significant.
The table below maps notable changes. Identifiers are not stable across editions, so teams should always cite the edition year together with the entry number in reports and tickets.
| 2023 entry | 2025 treatment | Reason for change |
|---|---|---|
| Insecure output handling | Renamed improper output handling (LLM05) | Broader wording for all downstream sinks |
| Training data poisoning | Expanded to data and model poisoning (LLM04) | Covers fine-tuning and embedded backdoors |
| Model denial of service | Folded into unbounded consumption (LLM10) | Adds cost abuse and denial of wallet |
| Model theft | Folded into unbounded consumption (LLM10) | Extraction often relies on high query volume |
| Insecure plugin design | Absorbed into excessive agency and others | Plugins evolved into general tool calling |
| Overreliance | Reframed as misinformation (LLM09) | Emphasis on hallucination and its effects |
Two entries were new in 2025: system prompt leakage and vector and embedding weaknesses. The working group has also produced companion material on agentic AI threats, because autonomous agents chain many model calls and tools, which multiplies the impact of every item on the main list.
Mechanisms behind the most frequent risks
Most entries share one root cause: an LLM cannot reliably separate instructions from data. Every token in its context window can influence behavior, so text from a web page, an email or a retrieved document can act as a command. This is why prompt injection holds first place in both editions and why indirect prompt injection through retrieved content is treated as part of LLM01.
A second recurring mechanism is implicit trust in model output. When generated text is rendered as HTML, executed as SQL or passed to a shell without validation, classic flaws such as cross-site scripting, injection and remote code execution return through a new channel. A third mechanism is privilege concentration: an agent holding broad credentials turns a single manipulated response into real actions, which is the essence of excessive agency.
- Untrusted input to the model drives prompt injection, poisoning and embedding attacks.
- Untrusted output from the model drives improper output handling and misinformation.
- Excess capability around the model drives excessive agency and unbounded consumption.
- Secrets placed near the model drive sensitive information disclosure and system prompt leakage.
Applying the list in design, testing and governance
The list is most effective when it is mapped to concrete controls at each stage of the delivery lifecycle. Architects use it during threat modeling, testers convert entries into test cases for AI red teaming, and governance teams reference it in policies and vendor questionnaires. The decision table below links typical situations to recommended actions.
| Situation | Relevant entries | Recommended control |
|---|---|---|
| Chatbot reads external web pages | LLM01, LLM05 | Isolate retrieved text and sanitize all output |
| Agent can send email or pay invoices | LLM06, LLM01 | Least privilege and human approval for actions |
| RAG over internal documents | LLM08, LLM02 | Enforce document-level access control at retrieval |
| Third-party model or adapter adopted | LLM03, LLM04 | Verify provenance, signatures and model cards |
| Public API with metered model calls | LLM10 | Rate limits, quotas and spending alerts |
| Answers used for medical or legal advice | LLM09 | Grounding, citations to source data and human review |
Practitioners often combine the list with AI guardrails, content filters and conventional controls such as input validation, output encoding, secret management and logging. No single control closes an entry completely, so defense in depth remains the guiding principle.
Limitations and common misreadings
The ranking reflects expert consensus and practitioner surveys, not a statistically measured frequency of incidents. Its order indicates perceived importance, and a lower position does not mean a risk is harmless in a particular system. The list is also not a compliance standard: passing a checklist review does not certify that an application is secure.
| Misreading | Why it is wrong | Better practice |
|---|---|---|
| A good system prompt prevents injection | Instructions in the prompt can be overridden | Treat model output as untrusted regardless |
| The list covers model alignment | It targets application security risks | Pair it with alignment and safety evaluations |
| Guardrail products solve all ten | Filters miss novel phrasing and encodings | Layer architectural controls with filters |
| Entry numbers are permanent | Numbering changed between editions | Cite edition year and entry name |
Finally, the list deliberately stays technology-neutral. It does not prescribe specific products or model providers, so each organization must translate the guidance into controls that fit its architecture, data sensitivity and regulatory obligations.
A non-technical view
Think of the OWASP Top 10 for LLM Applications as a list of the ten most common ways AI-powered software gets into trouble. Security experts from many companies compared their experiences and agreed on which problems matter most.
Many of these problems come from one simple fact: an AI model treats everything it reads as potential instructions. A hidden sentence on a web page can therefore trick a helpful assistant into doing something its owner never intended. The list helps builders ask the right questions before launch, such as what the AI is allowed to do, what it can see, and who checks its answers.
A concrete example
A mid-sized insurance company plans to launch an AI assistant that reads customer emails, looks up policy records and drafts replies. Before release, its security team runs a two-day threat modeling workshop and walks through each entry of the 2025 edition. They discover that the assistant could read any policy in the database (LLM08 and LLM02) and send replies without review (LLM06).
The team restricts retrieval to the policies of the sender, removes the direct sending capability and adds a human approval step. During testing, a red teamer hides an instruction in an email asking the assistant to forward another customer’s records. The attack now fails, because the assistant has no access to that data and cannot act without approval. The company also sets a monthly spending cap on model calls to address unbounded consumption.