Overview

The 2025 edition of the OWASP Top 10 for Large Language Model Applications. Pick a risk — every page shares this nav, language and theme.

10Risk categories
2025Edition
106Cited sources

What's New in 2025

A note from the project leads and the key changes in this edition.

LLM01: Prompt Injection

User prompts alter an LLM's behavior or output in unintended ways, even via content imperceptible to humans.

LLM02: Sensitive Information Disclosure

LLMs can expose PII, proprietary algorithms, or confidential business data through their outputs, causing privacy and IP breaches.

LLM03: Supply Chain

LLM supply chains are vulnerable to tampering and poisoning across training data, third-party models, adapters, and deployment platforms.

LLM04: Data and Model Poisoning

Manipulating training, fine-tuning, or embedding data introduces vulnerabilities, backdoors, or biases that compromise model integrity.

LLM05: Improper Output Handling

Insufficient validation and sanitization of LLM outputs before downstream use can enable XSS, SQL injection, SSRF, or remote code execution.

LLM06: Excessive Agency

Excessive functionality, permissions, or autonomy granted to LLM systems enables damaging actions from unexpected or manipulated LLM outputs.

LLM07: System Prompt Leakage

System prompts may inadvertently expose sensitive data or controls that attackers can extract and exploit to facilitate further attacks.

LLM08: Vector and Embedding Weaknesses

Weaknesses in how RAG vectors and embeddings are generated, stored, or retrieved can be exploited to inject harmful content, manipulate outputs, or leak data.

LLM09: Misinformation

LLMs producing false or misleading yet credible-seeming content, often from hallucination, leading to breaches, reputational and legal harm.

LLM10: Unbounded Consumption

Excessive, uncontrolled LLM inference enables denial of service, economic loss, model theft, and service degradation.