OWASP Top 10 for LLM applications

The OWASP Top 10 for LLM applications is a list of the ten most critical security risks in apps built on large language models (LLMs). OWASP, the Open Worldwide Application Security Project, publishes it.

The 2026 list

IDRiskWhat it is
LLM01Prompt InjectionText that overrides the model's instructions
LLM02Sensitive Information DisclosurePersonal or confidential data in outputs
LLM03Excessive AgencyMore tools, permissions or autonomy than needed
LLM04Supply ChainCompromised third-party models, data or code
LLM05Data and Model PoisoningTampered training or embedding data
LLM06Unbounded ConsumptionUncontrolled use that drains budget or capacity
LLM07MisinformationFalse output that appears credible
LLM08Hidden Context ExposureThe system prompt and other hidden context extracted
LLM09Vector and Embedding WeaknessesFlaws in embedding storage and retrieval
LLM10Improper Output HandlingModel output passed on without validation

Editions

OWASP revises the list, and the IDs move with it. The 2026 edition came out in August 2026. Against the 2025 list, Excessive Agency rose from sixth to third and Unbounded Consumption from tenth to sixth, and System Prompt Leakage widened into Hidden Context Exposure. So cite the year with an ID. Every edition is on the OWASP project's site.

How to use it

Use it as a checklist. For each row, name the control you have and the test that proves it.

With White Circle

Control enforces your policies on every input, output and tool call. Coverage includes prompt injection, leaks of confidential data, hallucinations, tool abuse and token drain.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL