OWASP GenAI LLM Top 10 (2026)
Published by OWASP on August 4, 2026
The ten risks that matter most when you put a large language model in front of real users and real data — what each one means in practice, and how an assistant platform can be designed to reduce it.
What changed in 2026
The 2026 edition reordered eight of the ten entries and renamed one. Excessive Agency climbed to #3, Unbounded Consumption jumped from #10 to #6, Misinformation rose to #7, and “System Prompt Leakage” was broadened and renamed “Hidden Context Exposure.” The ranking blends 6,639 documented real-world incidents with practitioner consensus.
A shared responsibility
Security here is shared. Some risks are the platform’s to carry, some belong to the application embedding the assistant, and some sit with the model provider. Below, each entry is honest about which is which — and where a control is defense-in-depth rather than a guarantee.
- LLM01Prompt InjectionDefense in depth
Crafted input — typed directly, or hidden inside content the model later reads — overrides your instructions and makes the model misbehave.
No change — #1 again in 2026 - LLM02Sensitive Information DisclosureStrong by design
The model reveals data it shouldn’t — PII, secrets, or another user’s records — in its answer or by over-fetching from connected systems.
No change — #2 again in 2026 - LLM03Excessive AgencyStrong by design
The model is granted more autonomy, permission, or functionality than the use case needs — so it can take consequential actions, especially when manipulated.
Climbed from #6 to #3 in 2026 - LLM04Supply ChainShared responsibility
Third-party components — models, libraries, datasets, plugins — can be compromised, backdoored, or vulnerable, and the weakness propagates to you.
Moved from #3 to #4 in 2026 - LLM05Data and Model PoisoningLow exposure by design
Training data or a model is corrupted — before or during training — so it behaves maliciously, carries a backdoor, or is subtly biased.
Moved from #4 to #5 in 2026 - LLM06Unbounded ConsumptionShared responsibility
Without limits on how often and how heavily a model is invoked, attackers can run up cost (“denial of wallet”), exhaust resources, or extract the model by querying it at volume.
Jumped from #10 to #6 in 2026 - LLM07MisinformationShared responsibility
The model produces confident but false or fabricated information that users may trust and act on.
Rose from #9 to #7 in 2026 - LLM08Hidden Context ExposureStrong by design
The invisible layer you place in front of the model — system instructions, tool definitions, retrieved documents, memory, configuration — leaks to users or attackers.
Renamed and broadened from “System Prompt Leakage” - LLM09Vector and Embedding WeaknessesNot applicable by design
Security gaps specific to retrieval systems — reconstructing source text from embeddings, cross-tenant leakage in a shared vector store, or poisoned embeddings.
Moved from #8 to #9 in 2026 - LLM10Improper Output HandlingStrong by design
Model output is trusted and passed unvalidated into something that executes it — rendered as HTML (XSS), run as SQL, or sent to a downstream system.
Fell from #5 to #10 in 2026