OWASP Top 10 for LLM Applications 2026: what changed, and where approval helps
All ten entries of the 2026 list, what moved since 2025 (LLM06 is now Unbounded Consumption, Excessive Agency is LLM03), and where a person's approval really helps.
· 8 min read
The OWASP GenAI Security Project published the Top 10 for LLM Applications 2026 on 3 August 2026. It is the first edition checked against evidence: OWASP gathered 7,714 real incidents from public vulnerability databases and an AI harm database, sorted the 6,639 with enough detail, and weighed that record at one quarter against the practitioners’ vote at three quarters.
The order moved more than in earlier years. Excessive Agency rose to third, which OWASP calls “the most consequential move on the list”. Unbounded Consumption rose four places to sixth. System Prompt Leakage became Hidden Context Exposure, and Improper Output Handling fell from fifth to tenth.
If you are looking for LLM06
The numbers changed, so check which edition a reference means. LLM06:2025 is Excessive Agency, which is now LLM03:2026. Our Excessive Agency article covers both. LLM06:2026 is Unbounded Consumption, which was LLM10:2025, and it has its own article. A policy, audit finding or tool that says only “LLM06” probably means the 2025 list.
What moved from 2025
| Option | In the 2025 list | Change |
|---|---|---|
| LLM01:2026 Prompt Injection | In the 2025 listLLM01:2025 Prompt Injection | ChangeSame place |
| LLM02:2026 Sensitive Information Disclosure | In the 2025 listLLM02:2025 Sensitive Information Disclosure | ChangeSame place |
| LLM03:2026 Excessive Agency | In the 2025 listLLM06:2025 Excessive Agency | ChangeUp 3 |
| LLM04:2026 Supply Chain | In the 2025 listLLM03:2025 Supply Chain | ChangeDown 1 |
| LLM05:2026 Data and Model Poisoning | In the 2025 listLLM04:2025 Data and Model Poisoning | ChangeDown 1 |
| LLM06:2026 Unbounded Consumption | In the 2025 listLLM10:2025 Unbounded Consumption | ChangeUp 4 |
| LLM07:2026 Misinformation | In the 2025 listLLM09:2025 Misinformation | ChangeUp 2 |
| LLM08:2026 Hidden Context Exposure | In the 2025 listLLM07:2025 System Prompt Leakage | ChangeRenamed, down 1 |
| LLM09:2026 Vector and Embedding Weaknesses | In the 2025 listLLM08:2025 Vector and Embedding Weaknesses | ChangeDown 1 |
| LLM10:2026 Improper Output Handling | In the 2025 listLLM05:2025 Improper Output Handling | ChangeDown 5 |
Several entries also grew. Prompt Injection now covers instructions hidden in images and audio, Supply Chain covers model artifacts that are not what they claim to be, Data and Model Poisoning takes in fine-tuning subversion, and Improper Output Handling covers insecure code that assistants generate.
The ten entries, and whether approval helps
The middle column is our reading, not OWASP’s: how much a person’s approval, or a standing rule that answers for them, does against each entry when the model drives an agent. The last column is a short version of OWASP’s own mitigations.
| Option | What goes wrong | Does approval help? | What helps most |
|---|---|---|---|
| LLM01:2026 Prompt Injection | What goes wrongText the model reads, from a user, a web page, an email, a tool result or its own memory, changes what it does. More on LLM01 | Does approval help?Partly. It cannot stop the injection, but an injected agent still has to ask before the action that does the harm. OWASP lists human confirmation as mitigation 7. | What helps mostAssume it will succeed: credentials in code, a policy check before privileged calls, person-approval for any agent that reads untrusted input, holds private data and can act. |
| LLM02:2026 Sensitive Information Disclosure | What goes wrongConfidential data comes out through an answer, a tool argument, a log or a reasoning trace to someone who should not see it. | Does approval help?Little. Most leaks happen inside an answer. A gate helps only where the leak is an action, such as an email or upload going out. | What helps mostClassify data before it reaches the model, redact in every output channel, scope retrieval to the person asking. |
| LLM03:2026 Excessive Agency | What goes wrongThe model can do damaging things because it has more tools, permissions or autonomy than the job needs. More on LLM03 | Does approval help?Directly. OWASP's mitigation 6 is to require user approval of high-impact actions, with low-consequence ones allowed to approve themselves. | What helps mostFewer tools, narrower permissions, a person for high-impact actions, limits on the cumulative value an agent may move. |
| LLM04:2026 Supply Chain | What goes wrongA third-party model, dataset, adapter or serving platform is tampered with or is not what it claims to be. | Does approval help?Little. A tampered model or package is a build problem. A gate only meets it once the bad component tries an action that asks. | What helps mostSigned and verified models and adapters, an inventory of components, vetted suppliers and conversion pipelines. |
| LLM05:2026 Data and Model Poisoning | What goes wrongTraining, fine-tuning or retrieval data is manipulated so the model behaves badly, often while looking fine. | Does approval help?Little. Poisoning happens before the model runs. Approval limits what a poisoned model can do, not the poisoning. | What helps mostTrack where training and retrieval data came from, validate it, test models for backdoors, review retraining loops. |
| LLM06:2026 Unbounded Consumption | What goes wrongUncontrolled use runs up the bill, takes the service down or lets someone copy the model: floods, reasoning loops, tool-call fan-out. More on LLM06 | Does approval help?Partly. Tokens are capped in your model gateway. An approval service caps the actions that cost money and stops every agent at once. | What helps mostToken-aware rate limits, hard spending caps that stop work, agent circuit breakers with step and cost limits. |
| LLM07:2026 Misinformation | What goes wrongConfident but wrong output is trusted and acted on, by a person, a workflow or an agent's next tool call. | Does approval help?Partly. OWASP's mitigation 5 asks for approval workflows, and its example is an agent approving a refund the policy forbids. | What helps mostGround claims before acting, separate generating from executing, approval workflows for high-impact actions. |
| LLM08:2026 Hidden Context Exposure | What goes wrongSomeone extracts the system prompt, tool schemas or other hidden context, and with it secrets or the rules to get around. | Does approval help?Little. A leaked prompt is a design problem. Keeping the policy in code outside the model is what makes a leak harmless. | What helps mostKeep secrets and authorisation logic out of prompts, assume the system prompt and tool schemas will be read. |
| LLM09:2026 Vector and Embedding Weaknesses | What goes wrongSimilarity search shows the model the wrong content: another tenant's data, planted documents, or text that wins by geometry. | Does approval help?Little. This is about what the model is shown. Approval only meets it later, at an action the retrieved text led to. | What helps mostPermission-aware retrieval, separate indexes per tenant, validation of what goes into the vector store. |
| LLM10:2026 Improper Output Handling | What goes wrongModel output goes into a browser, a query, a shell or an API without being checked, and becomes XSS, SQL injection or code execution. | Does approval help?Little. Escaping and parameterising are the fix. Approving each shell command is possible but tiring. | What helps mostTreat model output as untrusted input: encode it, parameterise queries, sandbox generated code. |
Where this list ends and the agentic one begins
OWASP draws the line itself. This list “owns the risk when the model is a component inside your application”. Once the model has tools it can call, memory it keeps between sessions and consequences downstream, “the risk moves to the OWASP Agentic Top 10”, and OWASP advises reading both, “because neither one covers that ground alone.” Our overview of the Top 10 for Agentic Applications does the same exercise for ASI01 to ASI10.
The project leads’ letter sets the tone for both: “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.”
What an approval step covers, and what it does not
One entry lands on the approval step. LLM03:2026 Excessive Agency asks for a person to approve high-impact actions, with a graduated policy that lets easily reversible ones run on their own.
Three more are helped part of the way. Prompt injection (LLM01) and misinformation (LLM07) both end in an action the model should not have taken, and both entries list human confirmation of high-impact actions among their mitigations. Unbounded Consumption (LLM06) needs hard caps that stop work, which an approval service gives you for actions that cost money, next to the token limits in your model gateway.
The other six are mostly solved elsewhere: classification and redaction for disclosure, signing and provenance for the supply chain and poisoning, keeping secrets out of prompts, permission-aware retrieval, and treating model output as untrusted input.
An approval is only a control if it sits in the code that performs the action and fails closed: pending, denied, expired or unreachable all mean the action does not happen. An instruction in the prompt to “ask first” is still the model deciding, and LLM08 assumes that prompt will be read.
Where to start
- LLM01:2026 Prompt Injection for agents: the Rule of Two in code, so an agent that has read untrusted text asks before it acts.
- LLM03:2026 Excessive Agency: the three root causes and what an approval gate fixes.
- LLM06:2026 Unbounded Consumption: circuit breakers, spending caps that stop work, and a pause for every agent at once.
To turn the list into decisions for your own agent, paste its tools into the tool risk checker, then write down who decides what with the approval policy generator. Both run in the browser without an account.