OWASP Top 10 for LLM Applications 2026: All 10 Risks Explained
Updated on November 19, 20266 minutes read
The OWASP Top 10 for LLM Applications is the OWASP GenAI Security Project's list of the most critical security risks for apps built on large language models (LLMs). The 2026 edition was published in August 2026 (its resource page is dated 3 August) and announced on 1 September 2026. This guide lists all ten entries with their official IDs, explains each one briefly, gives one practical control per entry and summarizes what changed from 2025.
The project leads' opening letter sets the tone: "Stop trying to build a model that cannot be fooled. Build the system around it, so that when the model is fooled, and it will be, nothing important breaks."
How the 2026 List Was Built
Earlier editions relied on practitioner judgment. For 2026, OWASP also tested the community vote against a corpus of 7,714 real incidents from public vulnerability databases and an AI-harm database, 6,639 of which had enough detail to classify. The vote carries three-quarters of the weight and the incident data one quarter. The two did not always agree: ranked by raw incident counts, prompt injection falls out of the top 10, which OWASP attributes to a "defense effect", since teams fight it hard and fewer clean exploits reach public databases. Misinformation went the other way: voters placed it near the bottom, the incident record near the top, and the evidence pulled it up to seventh.
The 2026 List at a Glance
| 2026 entry | Position in 2025 |
|---|---|
| LLM01:2026 Prompt Injection | 1 |
| LLM02:2026 Sensitive Information Disclosure | 2 |
| LLM03:2026 Excessive Agency | 6 |
| LLM04:2026 Supply Chain | 3 |
| LLM05:2026 Data and Model Poisoning | 4 |
| LLM06:2026 Unbounded Consumption | 10 |
| LLM07:2026 Misinformation | 9 |
| LLM08:2026 Hidden Context Exposure | 7 (as System Prompt Leakage) |
| LLM09:2026 Vector and Embedding Weaknesses | 8 |
| LLM10:2026 Improper Output Handling | 5 |
The 2025 positions come from OWASP's 2025 list.
LLM01:2026 Prompt Injection
Input that changes the model's behavior in ways the developer did not intend, whether it comes from a user, a document, a tool result, an image or memory. The entry now covers cross-modal attacks that hide instructions in images or audio. Control: keep credentials and state-changing actions in application code and give the model least privilege per operation. Our guide to prompt injection covers this entry in depth.
LLM02:2026 Sensitive Information Disclosure
Confidential or regulated data leaking through any channel: the answer, tool-call arguments, reasoning traces, retrieved chunks, logs or embeddings. OWASP names oversharing upstream, such as unscoped drives feeding a retrieval-augmented generation (RAG) index, as one of two failures behind most incidents. Control: authorize before retrieval, enforcing document- and chunk-level access inside the index query.
LLM03:2026 Excessive Agency
Damaging actions taken in response to unexpected, ambiguous or manipulated model output. The root causes are excessive functionality, permissions or autonomy. In agentic systems, OWASP says it can show up as tool misuse, identity and privilege abuse or cascading failures. Control: minimize tools and their permissions (an email-summary tool only needs to read mail) and require human approval for high-impact actions.
LLM04:2026 Supply Chain
Risks in third-party models, datasets, adapters such as LoRA, conversion pipelines and packages, including "slopsquatting", where attackers register package names that coding assistants hallucinate. Control: pin model artifacts by immutable digest rather than a mutable tag, verify signatures and check that AI-suggested dependencies exist before installing them.
LLM05:2026 Data and Model Poisoning
Manipulated data or model artifacts that embed backdoors or bias during pre-training, fine-tuning, embedding or distribution. Poisoning can be cheap: Souly et al. (2025) found that 250 poisoned documents compromised models from 600M to 13B parameters regardless of dataset size. Control: track data and model lineage, version your datasets and treat chat templates, tokenizer configs and adapters as code that must be signed and verified.
LLM06:2026 Unbounded Consumption
Uncontrolled inference that causes outages, runaway costs ("denial of wallet") or model theft through extraction. Reasoning models, multimodal inputs and agent tool loops make it worse. Control: hard spending caps that halt inference, plus agent circuit breakers with step, recursion, time and per-run cost limits.
LLM07:2026 Misinformation
Incorrect or unsupported output that looks credible enough to be acted on. In agentic systems a wrong answer becomes a wrong action, such as an agent reporting that a backup ran when it never did. Control: a claim-check-act pattern that separates generation from execution and verifies claims and preconditions before acting.
LLM08:2026 Hidden Context Exposure
Extraction or reconstruction of hidden context: system prompts, developer instructions, retrieved policy text and tool schemas. OWASP advises assuming that hidden context is discoverable. Control: keep credentials and secrets out of it, and enforce authorization in deterministic code outside the model.
LLM09:2026 Vector and Embedding Weaknesses
Attacks on the embedding layer behind RAG, agent memory and semantic caches. In OWASP's words, "poisoning makes the system wrong, inversion makes it leak, jamming makes it silent, and access-control failure makes it indiscriminate." Control: enforce tenant scoping inside the index query, not as a post-retrieval filter, and keep content of different trust levels in separate indexes.
LLM10:2026 Improper Output Handling
Passing model output downstream without validation, which can lead to XSS, SQL injection, SSRF or remote code execution. It now also covers insecure code that assistants generate at scale. Control: treat model output like user input: encode it for its context, use parameterized queries and disable auto-rendering of Markdown images and link previews.
What Changed From the 2025 Edition
OWASP says the order "moved more than in past years". The main changes:
- Excessive Agency climbed from sixth to third, which OWASP calls "the most consequential move on the list".
- Unbounded Consumption rose four places, and Improper Output Handling fell the furthest, from fifth to tenth.
- Prompt Injection and Sensitive Information Disclosure kept the top two places.
- System Prompt Leakage became Hidden Context Exposure, a broader entry.
- Existing entries absorbed newer risks instead of new categories being added: cross-modal injection, model artifacts that are not what they claim to be, fine-tuning subversion and insecure generated code.
- Framework mappings (NIST, MITRE ATLAS, CWE and others) moved into one appendix.
OWASP also draws a clearer boundary. This list covers the model as a component of your application. Once the model becomes an actor with tools, memory and downstream consequences, OWASP says to pair it with the OWASP Top 10 for Agentic Applications, which our guide on how to secure AI agents walks through.
How to Use the List in Your Team
OWASP's advice is to "work all ten, start at the top". In practice:
- Walk through the ten entries when you design or review an LLM feature, and note where it is exposed.
- Turn each control into a ticket with an owner, and re-test after fixes.
- Tag findings with the entry IDs so developers, security engineers and product owners share one vocabulary.
- Only test systems you own or have written permission to test, and check that your model provider's terms allow security testing.
The capstone of the self-paced AI Security: Securing LLM Apps & AI Agents course, Attack, Fix and Re-Test an AI Agent, ends with a security report that maps each finding to the OWASP lists.
Where to Go Next
Start with the entry at the top: our guide on what prompt injection is explains direct and indirect attacks and layered defenses. If your app retrieves documents, look closely at LLM02 and LLM09, which cover retrieval permissions and the embedding layer; AI Engineer: Introduction to RAG & Search is a related self-paced course.
One structured way to build these skills is the AI Security: Securing LLM Apps & AI Agents course: self-paced, Intermediate level and 40 hours long, for developers, AI engineers and security professionals who build or test AI features.

Learn technical skills online with Code Labs Academy
Join our supportive community, unlock your potential, and embark on a rewarding career path.
Apply Now