The OWASP Top 10 for LLM Applications
OWASP's Top 10 for LLM Applications is the closest thing this field has to a shared vocabulary — a ranked list of the vulnerability categories that show up again and again in real LLM-based systems (OWASP, 2025). Use it as a checklist for what a review needs to cover, not a compliance box to tick once.
The list¶
- Prompt Injection — user input alters the model's behavior in ways the system designer didn't intend, overriding instructions or extracting data it shouldn't share.
- Sensitive Information Disclosure — the model exposes private, proprietary, or otherwise sensitive information through its outputs.
- Supply Chain — vulnerabilities inherited from third-party models, training data, fine-tuning services, or plugins the system depends on.
- Data and Model Poisoning — pre-training, fine-tuning, or embedding data is manipulated to introduce vulnerabilities, backdoors, or bias.
- Improper Output Handling — insufficient validation or sanitization of what the model produces before it's used downstream (rendered in a browser, executed as code, passed to another system).
- Excessive Agency — an agentic system is granted more autonomy, permissions, or tool access than its actual task requires.
- System Prompt Leakage — the instructions meant to stay hidden from the user get exposed, handing an attacker the system's own playbook.
- Vector and Embedding Weaknesses — security gaps specific to retrieval-augmented systems, where the vector store or embedding pipeline becomes the attack surface.
- Misinformation — the model generates false or misleading content that's presented with unwarranted confidence.
- Unbounded Consumption — the system allows unrestricted resource use, covering denial-of-service, cost exploitation, and unauthorized model replication.
What's changed from earlier versions¶
This list evolves as real-world incidents surface new patterns. One thing worth knowing if you've used an older version of this list: model theft is now folded into Unbounded Consumption rather than standing alone — OWASP's own description explicitly covers stealing IP by cloning a model's behavior through API extraction (OWASP, 2025). Treat this as a living list — check for the current version before you scope a review, not this page's snapshot of it.
Sources¶
- OWASP (GenAI Security Project) — OWASP Top 10 for LLM Applications, 2025. LLM01:2025 Prompt Injection; LLM02:2025 Sensitive Information Disclosure; LLM03:2025 Supply Chain; LLM04:2025 Data and Model Poisoning; LLM05:2025 Improper Output Handling; LLM06:2025 Excessive Agency; LLM07:2025 System Prompt Leakage; LLM08:2025 Vector and Embedding Weaknesses; LLM09:2025 Misinformation; LLM10:2025 Unbounded Consumption. View source · verified 2026-07-02 · primary
- OWASP (GenAI Security Project) — LLM10:2025 Unbounded Consumption, 2025. Attacks designed to disrupt service, deplete the target's financial resources, or even steal intellectual property by cloning a model's behavior all depend on a common class of security vulnerability... Model Extraction via API: Attackers may query the model API using carefully crafted inputs and prompt injection techniques to collect sufficient outputs to replicate a partial model or create a shadow model. View source · verified 2026-07-02 · primary