Guides

OWASP Top 10 for LLM Applications Explained (2026) 

8
Mins Read

Table of Contents

Still relying on traditional DLP for AI?
There's a better way.

Semantic Understanding

Real Time Inline Action

Dynamic Policy Engine

Get A Free POC

Trusted by 55+ regulated organizations

Key takeaways

  • The 2026 OWASP Top 10 for LLM applications keeps all ten risks but reorders eight of them, and excessive agency, the risk of an AI system holding more power than its task needs, climbs to third.
  • Two risks sit close to everyday work for a small team: Sensitive Information Disclosure (what employees send in) and Excessive Agency (what connected agents can do), which is where AI DLP applies.
  • The other eight mostly belong to whoever builds or hosts the model, so the practical move for most teams is to ask vendors about them.

In August 2026 OWASP reordered eight of the ten entries in its Top 10 for LLM Applications, and the biggest climber was Excessive Agency, the risk of giving an AI system more power to act than its job needs. The change matters because, for the first time, the ranking drew on real incident data alongside expert votes, and the incidents point at agents and connected tools.

If you run IT or security as a small team in the USA or India, the OWASP Top 10 for LLM is worth reading, even though it was written for people building AI applications. Below we go through all ten risks in the 2026 order, and for each one we say what it means for employees who use AI tools at work.

AI-Native DLP and the Future of Enterprise AI Security

An executive whitepaper on how AI-Native DLP differs from legacy DLP, and what it means for enterprise security strategy.

Download Whitepaper

‍

OWASP Top 10 for LLM, 2026

Eight of ten risks moved. Two sit with the people using AI.

01 Prompt Injection
06 Unbounded Consumption
02 Sensitive Information Disclosure
07 Misinformation
03 Excessive Agency
08 Hidden Context Exposure
04 Supply Chain
09 Vector and Embedding Weaknesses
05 Data and Model Poisoning
10 Improper Output Handling
Where employees decide

The OWASP Top 10 for LLM and what changed in 2026

The OWASP Top 10 for LLM, also called the OWASP LLM Top 10, is a ranked list of the most serious security risks in applications built on large language models. The OWASP GenAI Security Project maintains it and publishes the current 2026 list with a ranking that is now 75% practitioner vote and 25% evidence from real incidents.

No entry was dropped, and none was added, so the headline is "movement." If you are still searching for the 2025 version, here is how the two compare.
‍

2026 Risk 2025 position
LLM01 Prompt Injection 1
LLM02 Sensitive Information Disclosure 2
LLM03 Excessive Agency 6
LLM04 Supply Chain 3
LLM05 Data and Model Poisoning 4
LLM06 Unbounded Consumption 10
LLM07 Misinformation 9
LLM08 Hidden Context Exposure (renamed from System Prompt Leakage) 7
LLM09 Vector and Embedding Weaknesses 8
LLM10 Improper Output Handling 5


1. LLM01: Prompt Injection

Prompt injection happens when instructions hidden in text the model reads, such as an email, a web page, or an uploaded file, change what the model does. The 2026 entry also covers attacks hidden in images and audio. No one has to type anything malicious into a chat window for this to work.

For employees using AI: a summary of an outside document can quietly carry instructions, so treat anything an AI tool does after reading untrusted content with the same caution as an unknown attachment.

2. LLM02: Sensitive Information Disclosure

Sensitive information disclosure covers everything a model reveals that it should not: data memorized in training, confidential content pulled into a response, or a retrieval system that returns a document nobody meant to share. It also covers what goes in, since anything pasted into a prompt can come back out somewhere else. This is the risk our prompt security work is built around.

The wider idea of data exposure is the useful frame here, because most disclosure is an ordinary working day rather than an attack. Most of the time it starts without an attacker, with an ordinary employee pasting customer data, PII, source code, or API keys into a tool to save time. A rule that matches known patterns will catch a card number but miss a pasted paragraph of unreleased pricing. 

Manual policies can only address known behaviors, but AI interactions are dynamic and constantly evolving, so by the time a new data exposure risk is discovered, it may be too late to configure a policy to prevent it. Our explainer on data exfiltration shows how this plays out on a normal working day.

For employees using AI: assume anything you paste leaves your control, and ask for the policy in plain words so you know what is safe to share.

3. LLM03: Excessive Agency

Excessive agency is what happens when an AI system can act, not just answer, and holds more permission than its task requires. A model that can only read files is an exposure risk; one that can also send, delete, or edit them turns a bad output into a real action. OWASP breaks the cause into too much functionality, too many permissions, and too little human approval.

For employees using AI: every connection you give an assistant, whether to your inbox, drive, or code repository, is a permission, so grant the narrowest one that does the job.

This is also why we pay attention to agents and MCP connections. Our approach is that policy follows the action, so a connected tool is held to the same rules as a person typing.

4. LLM04: Supply Chain

Supply chain risk covers every outside component an LLM application depends on: pretrained models, datasets, plugins, and the libraries that serve them. A tampered model file or a malicious package suggested by a coding assistant can introduce a weakness long before anyone sees odd output.

For employees using AI: install only the plugins, extensions, and coding-assistant packages your team has reviewed, since a convenient add-on is also a new dependency.

5. LLM05: Data and Model Poisoning

Poisoning means corrupting the material a model learns from or retrieves from, at pretraining, fine-tuning, or retrieval time, so that harmful behavior is built in. The 2026 entry now also covers fine-tuning subversion. Unlike most bugs, it cannot be patched, because the fix is to find what was corrupted and revalidate or retrain.

For employees using AI: if your company fine-tunes a model or feeds it a shared knowledge base, who can edit that material matters as much as who can read it.

6. LLM06: Unbounded Consumption

Unbounded consumption describes a system that has no ceiling on how much it can be made to compute, process, or cost. The 2026 entry reaches beyond outages to cost attacks, sometimes called denial of wallet, in which small requests fan out into many paid calls.

For employees using AI: a runaway agent or script that calls a paid AI service in a loop can run up a bill quickly, so set usage limits on any account tied to a company card.

7. LLM07: Misinformation

Misinformation is output that is wrong but confident enough to be trusted and acted on. OWASP treats it as a system-level failure now, because model output increasingly drives tool calls, code, and decisions, not only reading material.

For employees using AI: check any AI answer before it reaches a customer, a contract, or a production system, particularly when no one else will review it.

8. LLM08: Hidden Context Exposure

Hidden context exposure is the 2026 name for what used to be system prompt leakage. It covers everything assembled into a model's context that users are not meant to see, including system instructions, retrieved policy text, tool definitions, and internal rules. OWASP's advice is to assume that this context will eventually be discoverable, so what it reveals matters more than whether it leaks.

For employees using AI, do not put secrets or internal rules into a prompt or an assistant's instructions, because they are not as private as they look.

9. LLM09: Vector and Embedding Weaknesses

Vector and embedding weaknesses sit wherever similarity search stands between a data source and a prompt, such as in retrieval-augmented generation, agent memory, and semantic caches. In a shared setup, weak access control at that layer can let one customer's documents surface for another through the way queries are matched.

For employees using AI: when you connect a document library to an assistant, check who can see what, since the assistant may return files that people were never meant to find.

10. LLM10: Improper Output Handling

Improper output handling is what happens when a model's response reaches another system without checks, for example, rendered as a web page, run as code, or passed into a database query. It fell from fifth to tenth in 2026, but not because it shrank: the fix is well understood, which is to treat model output like any other untrusted input.

For employees using AI: do not paste AI-written code or commands straight into production without reading them.

Where AI DLP fits against the OWASP list

Our view at Wald AI is that the OWASP list is written for people building LLM applications, while most small teams are consuming them, and consumers still own two of the biggest risks: what employees send in and what connected agents are allowed to do. We read every entry through the employee at the keyboard, which is why we are open about what AI DLP does not cover. Most of this list belongs to the teams that build and host models. 

Wald AI is a solution that uses multiple small language models (SLMs) on your device to spot sensitive data in prompts, understand what it means, and apply the right policies before it is shared with ChatGPT, Claude, or another AI tool, so it helps with the parts that happen at the keyboard. The table says plainly where we help and where we do not.

Risk Where Wald AI helps Where it does not
LLM01 Prompt Injection Keeps sensitive data in prompts and uploads from leaving Does not stop the injection itself
LLM02 Sensitive Information Disclosure Our core use: classifies what a prompt means and applies policy before it is shared Does not change what a model memorized in training
LLM03 Excessive Agency Applies policy to agent and MCP actions Does not set the permissions inside the agent's own app
LLM04–LLM10 Not our layer Model-level and application-level risks belong to the builder or host.

Every interaction is logged, including what was detected and what action was taken. These audit logs help your team support its compliance efforts, while Wald AI's zero data retention policy and SOC 2 Type II report provide additional assurance.

How a small team can use the OWASP Top 10 for LLM

You do not need to work through all ten at once.

  1. List the AI tools and agents your team actually uses. Include personal accounts and connected apps, since those are where LLM02 and LLM03 live.
  2. Pick the risks that apply to you as a user. For most small teams, that means LLM02, LLM03, and LLM07, and a plan for each.
  3. Ask vendors about the rest. For LLM04 to LLM10, a short written answer from each AI vendor tells you more than a self-assessment.

A 30-person company in Austin, for example, might discover that three people connect ChatGPT to a shared drive and one pastes support tickets into a personal account. That single finding covers LLM02 and LLM03 at once, and it turns a ten-item list into two concrete decisions: which connections stay, and which prompts need a policy.

Conclusion

The 2026 OWASP Top 10 for LLM moves the center of gravity toward agents and the data they can reach. For a small team, that means two jobs: control what employees send into AI tools and limit what connected assistants are allowed to do. 

The other eight risks are real, but they are mostly for the people who build and host the models, so ask them about it. If you want to see how contextual classification handles the first two on your own team's AI use, explore AI DLP or talk to our team.

‍

Frequently Asked Questions

What is the OWASP Top 10 for LLM applications?

The OWASP Top 10 for LLM applications is a ranked list of the ten most serious security risks in software built on large language models, maintained by the OWASP GenAI Security Project. The 2026 edition runs from Prompt Injection at number one to Improper Output Handling at number ten.

Is prompt injection still the top LLM risk in 2026?

Yes. Prompt Injection stays at number one in the 2026 OWASP Top 10 for LLM, and Sensitive Information Disclosure stays at number two. The biggest change is excessive agency rising to third place.

What is sensitive information disclosure (LLM02)?

Sensitive information disclosure is the risk that an AI system exposes data it should not, whether that data was memorized in training, pulled into a response, or pasted into a prompt by an employee. For small teams it most often starts with an ordinary paste into a chat tool.

Does the list apply if we only use ChatGPT and do not build apps?

Partly. Most entries are about building and hosting models, but Sensitive Information Disclosure and Excessive Agency apply to anyone who uses AI tools at work, because employees decide what goes in and which connections an assistant is given.‍

Still relying on traditional DLP for AI?
There's a better way.

Semantic Understanding

Real Time Inline Action

Dynamic Policy Engine

Get A Free POC

Trusted by 55+ regulated organizations