Last updated: November 16, 2026
The OWASP Top 10 for LLM Applications lists the most important security risks in apps built on large language models, led by prompt injection, sensitive information disclosure and supply chain weaknesses. OWASP’s newer Top 10 for Agentic Applications covers risks that appear when AI agents plan and act with tools, such as goal hijacking, tool misuse and identity abuse. Together they are the most widely used checklists for securing AI apps in 2027.
Below we explain each risk in plain language with the controls that help most. Always check the official OWASP GenAI Security Project pages for the latest wording, since the lists are updated as threats change.
| ID | Risk | What it means | Key controls |
|---|---|---|---|
| LLM01 | Prompt Injection | Inputs or hidden content override the model’s instructions | Treat content as untrusted, least privilege, detection tools |
| LLM02 | Sensitive Information Disclosure | The model reveals private or confidential data | Data minimization, output filtering, access control |
| LLM03 | Supply Chain | Risky models, datasets or plugins | Vet and scan models, pin versions |
| LLM04 | Data And Model Poisoning | Tampered training or fine-tuning data | Data provenance and validation |
| LLM05 | Improper Output Handling | Model output used unsafely downstream | Validate and encode outputs |
| LLM06 | Excessive Agency | The model can do too much through tools | Limit tools, scopes and require approvals |
| LLM07 | System Prompt Leakage | Secrets or logic exposed via the system prompt | Keep secrets out of prompts |
| LLM08 | Vector And Embedding Weaknesses | Attacks on RAG stores and embeddings | Access control on vector data, input checks |
| LLM09 | Misinformation | Confident but wrong output | Grounding, citations, human review |
| LLM10 | Unbounded Consumption | Abuse that drives up cost or causes outages | Rate limits, quotas, monitoring |
The agentic list focuses on what changes when AI can plan, remember and act. Its themes include:
For a step-by-step version, use our AI agent security checklist.
It is a community-built list of the ten most important security risks in applications that use large language models, maintained by the OWASP GenAI Security Project.
Prompt injection is ranked first, because inputs or hidden content can override a model’s instructions.
Yes. OWASP published a Top 10 for Agentic Applications focused on risks such as goal hijacking, tool misuse and identity abuse.
Use them as a checklist during design, testing and procurement, and ask vendors and testers to map their controls and findings to them.
AI Security And Cyber Threats hub, How to secure AI agents and MCP integrations, Top MCP security tools for 2027, AI red teaming vs penetration testing.