OWASP Top 10 for Large Language Models (LLMs) Applications

OWASP released an OWASP Top 10 for Large Language Models (LLMs) Applications on 16th October, 2023. The version of this document is 0.6. This document is prepared by more than 500 researchers working in the field.

The primary purpose for the creation of the OWASP Top 10 for LLM applications is to mitigate the security and safety issues in large implementations utilizing LLM.

OWASP Top 10 for Large Language Models (LLMs) Applications

LLM01	Prompt Injections	This vulnerability allows attackers to inject craftily input into LLM applications, providing the intended results of the attacker
LLM02	Insecure Output Handling	This vulnerability arises when the LLM application accepts output without verifying the malicious nature
LLM03	Training Data Poisoning	This vulnerability allows the poisoning of data by malicious tainted training data
LLM04	Model Denial of Service	Here, vulnerability arises because of security issues during the development of LLM applications
LLM05	Supply Chain Vulnerabilities	Here, vulnerability arises because of security issues while the development of LLM applications
LLM06	Sensitive Information Disclosure	Reveal sensitive information in responses provided by LLM model
LLM07	Insecure Plugins Design	Remote Code Execution because of untrusted inputs
LLM08	Excessive Agency	This vulnerability arises due to providing excessive permissions in providing responses.
LLM09	Overreliance	LLM application provides an output that may be invalid
LLM10	Model Theft	Leakage of proprietary LLM model

Click Here for the Checklist to perform an audit of AI/ML systems

LLM Application Data Flow

Demonstrate high-level architecture for a hypothetical large language model application

Reference

OWASP Top 10 for Large Language Models (LLMs) Applications