OWASP Top 10 for LLM Applications: Risks and Controls (2025 and 2026)
The OWASP Top 10 for LLM Applications is the reference list of the most critical security risks for applications built on large language models, published by the OWASP GenAI Security Project. It is the starting point most organisations use to secure chatbots, copilots, RAG systems and AI agents.
Key takeaways
- Prompt Injection and Sensitive Information Disclosure remain the top two risks.
- The 2026 edition moves Excessive Agency up to third place, reflecting the rise of AI agents.
- System Prompt Leakage has been broadened into Hidden Context Exposure, which also covers tool schemas, permissions and workflow rules.
- Most controls are architectural: least privilege for tools, output handling, isolation of untrusted content and monitoring.
The 2025 list at a glance
| ID | Risk | Key controls |
|---|---|---|
| LLM01 | Prompt Injection | Treat all inputs and retrieved content as untrusted; separate instructions from data; input and output guardrails; least privilege for tools. |
| LLM02 | Sensitive Information Disclosure | Minimise data in context; access control on retrieval; redaction; output scanning for secrets and personal data. |
| LLM03 | Supply Chain | Vet models, datasets, plugins and libraries; pin versions; verify provenance and integrity. |
| LLM04 | Data and Model Poisoning | Control and audit training, fine-tuning and RAG sources; detect anomalies; keep datasets versioned. |
| LLM05 | Improper Output Handling | Never pass model output directly to shells, SQL, browsers or APIs; validate and encode it like any user input. |
| LLM06 | Excessive Agency | Limit tools, permissions and autonomy; require human approval for sensitive actions; scope credentials per task. |
| LLM07 | System Prompt Leakage | Never put secrets or authorisation logic in prompts; enforce rules outside the model. |
| LLM08 | Vector and Embedding Weaknesses | Access control in vector stores; tenant isolation; validate what is indexed. |
| LLM09 | Misinformation | Grounding and citations; human review for high-impact outputs; clear user warnings. |
| LLM10 | Unbounded Consumption | Rate limits, quotas, maximum output and tool-loop limits; cost monitoring. |
What changed in the 2026 edition
The OWASP GenAI Security Project published the 2026 edition in September 2026. It is an evolution rather than a revolution: most risk categories are unchanged, while rankings and definitions shift to reflect agentic AI.
- Prompt Injection stays first, with a wider definition of the attack surface, and Sensitive Information Disclosure stays second.
- Excessive Agency rises to third place: agents that can act on systems are now a primary concern.
- System Prompt Leakage is broadened into Hidden Context Exposure, which also covers tool schemas, permission models and workflow rules that an attacker can extract.
- Mappings to NIST, MITRE ATLAS, CWE and the OWASP Top 10 for Agentic Applications are expanded.
For the exact numbering of every 2026 entry, refer to the official document on genai.owasp.org.
The three risks that matter most for AI agents
Prompt injection
An agent that reads e-mails, tickets or web pages will eventually read an attacker's instructions. The defence is not a better prompt but an architecture where untrusted content cannot trigger privileged actions on its own.
Excessive agency
Give each agent the smallest set of tools and the narrowest permissions it needs, use short-lived credentials, and put a human approval step in front of irreversible actions such as payments, deletions or external messages.
Hidden context exposure
Assume anything in the context window — system prompt, tool definitions, retrieved documents — can be revealed. Keep secrets and authorisation decisions outside the model.
Turning the Top 10 into a programme
- Inventory your LLM applications, agents, models and tools.
- Assess each one against the Top 10 and record the gaps.
- Test the gaps with red teaming (see our AI red teaming guide).
- Deploy runtime controls and monitoring, then re-test after every change.
Frequently asked questions
Who publishes the OWASP Top 10 for LLM Applications?
The OWASP GenAI Security Project, a community-driven OWASP initiative. The documents are free at genai.owasp.org.
What is the number one risk in the OWASP Top 10 for LLM?
Prompt Injection (LLM01), in both the 2025 and 2026 editions.
What is new in the 2026 OWASP Top 10 for LLM?
Excessive Agency moves up to third place, System Prompt Leakage is broadened into Hidden Context Exposure, Prompt Injection gets a wider definition, and mappings to NIST, MITRE ATLAS, CWE and the Agentic Top 10 are expanded.
Does the OWASP Top 10 for LLM apply to AI agents?
Yes, and OWASP also publishes a dedicated Top 10 for Agentic Applications. Agents make Prompt Injection and Excessive Agency especially critical because a manipulated agent can act on real systems.
Related guides
- AI Red Teaming — Method, Garak vs PyRIT, and how to test LLM applications and agents.
- MITRE ATLAS — Tactics and techniques for attacks on AI, applied to agents.
- AI Agent & MCP Security — Threats to agents and MCP servers, with a practical checklist.