Skip to content

The OWASP Top 10 for LLM Applications 2025, Explained

A plain-English guide to the OWASP Top 10 for LLM Applications: what each risk means, why it matters, and how to test your AI system against it.

Wahid IqbalBy , Head of ComplianceUpdated 7 min read

AI & LLM Penetration TestingFrom $4,500Fixed scope, free retest, no hourly billing.

Get my fixed price

As organizations rush LLM features into production, they are shipping an attack surface that traditional security testing does not cover. The OWASP Top 10 for LLM Applications, now in its 2025 edition from the OWASP GenAI Security Project, is the reference list of the ten most critical risks in that new surface, the shared language for the risks specific to generative AI. Here is what each risk means in plain terms, and how it shows up in real AI penetration testing.

Why LLM apps need their own Top 10

An LLM has no hard boundary between the developer’s instructions and an attacker’s input. Vendors train models to favor system instructions, but that is a tendency, not an enforcement mechanism: the UK’s National Cyber Security Centre notes that current models do not enforce a security boundary between instructions and data inside a prompt. To the model, it is all just text in the context window. Add tools the model can call and data sources it can read (see how integrations expand the LLM attack surface), and you get failure modes that no web application checklist, the OWASP Top 10 2025 included, anticipates. The OWASP LLM Top 10 names them.

The ten risks in the 2025 edition, LLM01 to LLM10

The list below follows the 2025 edition, the current one, and uses its identifiers.

LLM01: Prompt injection. Crafted input that manipulates the model into ignoring its instructions or guardrails. It comes in two forms: direct (the user types it) and indirect prompt injection (the model reads it from a document, web page, or email it ingests). This is the defining risk of LLM security.

LLM02: Sensitive information disclosure. The model reveals training data, system prompts, secrets, or another user’s context through its output. Often triggered by prompt injection or weak output filtering.

LLM03: Supply chain. Risks inherited from third-party models, datasets, plugins, and libraries, a poisoned or compromised component upstream becomes your problem downstream. See also our piece on AI vendor risk.

LLM04: Data and model poisoning. Manipulating training or fine-tuning data (or a RAG knowledge base) to plant backdoors or bias the model’s behavior toward the attacker’s goal.

LLM05: Improper output handling. Treating model output as trusted and passing it, unvalidated, into downstream systems, where it can drive cross-site scripting, SQL injection, or unsafe code execution. The model becomes an injection vector.

LLM06: Excessive agency. Giving the model too much autonomy, too many tools, or too broad permissions, so that a manipulated model can take real, damaging actions (send email, move money, modify records) rather than just produce text. Model Context Protocol servers are now a common way to hand a model those tools; our guide to MCP server security testing covers how to test them, and agentic AI security covers agents as a whole. Our write-up of recent MCP vulnerabilities has dated examples.

LLM07: System prompt leakage. Exposure of the system prompt, which often contains instructions, logic, or secrets the developer assumed were hidden. Attackers use it to understand and bypass your guardrails.

LLM08: Vector and embedding weaknesses. Flaws in the retrieval (RAG) layer, poisoning the vector store, retrieval abuse, or weak access control on the data sources the model draws from.

LLM09: Misinformation. The model produces confident, plausible, and wrong output that users trust and act on. A safety and reliability risk as much as a security one.

LLM10: Unbounded consumption. Resource abuse, from denial-of-service through expensive queries to model-extraction attacks that clone your model’s behavior through excessive querying. Stolen API keys turn it into a direct bill: see LLMjacking and stolen AI keys.

What changed in the 2025 OWASP LLM Top 10

The 2025 list replaced the 2023-24 list (version 1.1). Two entries are new, several moved, and Insecure Plugin Design no longer has its own entry. Each row adds an example finding and the main fix.

2025 ID and risk Versus 2023-24 Example finding Main mitigation
LLM01 Prompt injection Still first A support bot follows instructions hidden in an uploaded PDF Treat retrieved content as untrusted; human approval for high-risk actions
LLM02 Sensitive information disclosure Up from LLM06 One tenant’s data appears in another tenant’s answer Access control before retrieval; keep secrets out of prompts
LLM03 Supply chain Up from LLM05 A model pulled from a public hub with no integrity check Pin and verify models and packages; keep an inventory
LLM04 Data and model poisoning Was LLM03 Training Data Poisoning Any user can write to the knowledge base the model retrieves from Control what enters training and retrieval data
LLM05 Improper output handling Was LLM02 Insecure Output Handling Model output rendered as HTML runs script in the browser Encode and validate output like any untrusted input
LLM06 Excessive agency Up from LLM08 An assistant that only needs to read mail can also delete it Narrow tool permissions; confirm writes
LLM07 System prompt leakage New An extractable system prompt contains an API key Keep secrets and authorization logic out of the prompt
LLM08 Vector and embedding weaknesses New Retrieval has no per-user filter and crosses tenants Permission-aware retrieval; vet documents before indexing
LLM09 Misinformation In the slot Overreliance held The assistant invents a package name a developer then installs Ground answers in sources; review before acting
LLM10 Unbounded consumption Covers the former Model Denial of Service and Model Theft entries No per-user rate or spend limit on a costly endpoint Quotas, rate limits and spend alerts

How to test against it

Reading the list is not the same as knowing where you stand. Real AI and LLM penetration testing works through these risks against your actual system:

  • Attempts direct and indirect prompt injection against your production guardrails (LLM01)
  • Tries to extract the system prompt, secrets, and other users’ data (LLM02, LLM07)
  • Maps every tool the model can call and tests for excessive agency and unsafe output handling (LLM05, LLM06)
  • Probes the RAG pipeline for poisoning and retrieval abuse (LLM04, LLM08)
  • Checks rate, token and spend limits on model endpoints, and where models and packages come from (LLM10, LLM03)
  • And remembers that the AI feature still sits inside a normal app that needs conventional web application and API testing too

Findings in the report are mapped to the OWASP Top 10 for LLM Applications and MITRE ATLAS, so each one carries the identifier it falls under, the input that triggered it and the fix.

Use it as a testing agenda

The OWASP Top 10 for LLM Applications is most valuable not as a checklist to read once, but as the agenda for a real assessment. If you are shipping LLM features, work through these ten risks against your own system before an attacker or a customer does. See what an LLM penetration test costs: a single LLM feature starts at $4,500, fixed before testing, with a free retest. Scope an AI/ML assessment and we will show you which of the ten your system is actually exposed to.

Frequently asked questions

What is the latest version of the OWASP Top 10 for LLM Applications? The 2025 edition, published by the OWASP GenAI Security Project. It replaced the 2023-24 list, and its entries run from LLM01:2025 to LLM10:2025.

What is the difference between direct and indirect prompt injection? In a direct injection the user’s own prompt alters the model’s behavior. In an indirect injection the model picks up instructions from external content it reads, such as a web page, a file or an email.

Does a web application penetration test cover LLM risks? Not by itself. It covers the application and API around the model: authentication, access control, injection. Prompt injection, system prompt leakage, tool abuse and retrieval flaws need LLM-specific tests, and most AI features need both.

Wahid Iqbal

Written by

Wahid Iqbal, Head of Compliance

20+ years in the field. Focus: PCI DSS, ISO 27001 & ISO 22301. Part of the team that scopes, tests, and reports for our clients. Full profile →

All articlesAI/ML Pentesting

Find out what an attacker sees.

Tell us what to test and see your fixed price.

Prefer the full scoping questionnaire? 
Start the conversation

Get a quick proposal

We are online