OWASP AI Security & LLM Top 10 is a community‑driven security standard that identifies the most critical security risks affecting Large Language Model (LLM) applications and generative AI systems. It is part of the broader OWASP GenAI Security Project, a global initiative focused on securing AI and agentic systems.
The framework is updated regularly (latest major update: 2026 edition) and is built from thousands of real‑world AI security incidents, contributions from hundreds of AI security experts, and mappings to industry frameworks such as NIST, MITRE ATLAS, CWE, and OWASP’s own Agentic Applications Top 10.
A ranked list of the top 10 security risks in LLM‑powered applications. It provides:
Clear explanations of each risk
Real attack scenarios
Practical mitigations
Engineering patterns
Mappings to industry frameworks
It is not a regulation, but it has become the de facto global standard for securing LLM applications.
Based on the 2025–2026 editions, the major categories include:
Prompt Injection – attackers override system instructions
Sensitive Information Disclosure – leaking secrets or private data
Supply Chain Security – compromised datasets, models, or dependencies
Data & Model Poisoning – malicious training or fine‑tuning data
Improper Output Handling – trusting model output without validation
Excessive Agency – LLMs performing unsafe or unintended actions
System Prompt Leakage – revealing internal instructions 8–10. Vector Database Weaknesses, Misinformation, Resource Abuse – risks in RAG pipelines, hallucinations, and model misuse
These risks reflect real adversary behaviour, not theoretical concerns.
OWASP AI Security & LLM Top 10 is essential because:
LLMs introduce new attack surfaces (prompt injection, agentic behaviour, RAG exploitation).
Traditional AppSec frameworks do not cover AI‑specific risks.
Organisations need practical engineering guidance to secure AI systems.
It is widely used in security audits, AI governance, and Zero Trust for AI programmes.
The 2026 edition also aligns with the Agent Control Standard (ACS) for securing agentic AI systems.