Zpedia 

/ What Is the OWASP Top 10 for LLM Applications?

What Is the OWASP Top 10 for LLM Applications?

The OWASP Top 10 for LLM Applications ranks the ten most critical large language model security risks, which are: prompt injection, sensitive information disclosure, excessive agency, supply chain, data and model poisoning, unbounded consumption, misinformation, hidden context exposure, vector and embedding weaknesses, and improper output handling. 

 

The OWASP GenAI Security Project ranked prompt injection as the #1 LLM security risk in the August 2026 edition of the report.

Who Maintains the OWASP Top 10 for LLM Applications, and How Is It Used?

The OWASP Top 10 for LLMs is a community-organized list of the ten most critical security risks involved in building and deploying large language model applications. The list is maintained by the OWASP GenAI Security Project, which publishes the OWASP Top 10 for Agentic Applications. OWASP also publishes other Top 10 lists across different projects, including the Top 10 Web Application Risks.

OWASP originally launched the Top 10 for LLMs project in 2023, and released the 2026 edition in August 2026. Practitioners and security leaders use the Top 10 as a risk-awareness tool to help inform enterprise GenAI security practices. It’s not used as a certification or compliance standard.

What Changed in the OWASP Top 10 for LLMs Between 2025 and 2026?

The August 2026 update of the OWASP Top 10 for LLM Applications kept most of the same risk areas but reordered a few key risks. Prompt Injection and Sensitive Information Disclosure remained ranked #1 and #2, respectively. Excessive Agency rose three places from #6 to #3, and Unbounded Consumption rose to #6. Misinformation rose two places to #7. Improper Output Handling fell from #5 to #10. 

System Prompt Leakage was renamed Hidden Context Exposure in 2026, and Prompt Injection includes cross-modal attacks. 

The 2026 edition also updated its methodology by using two inputs: a community vote from practitioners that was weighted at 75%, and an analysis of 7,714 real-world incidents that carried the remaining 25%. The incident analysis dataset included events from public vulnerability databases and an AI harm database, and had enough details from 6,639 events for OWASP to include them in the analysis.

Practitioner opinion and the incident analysis data conflicted in two notable places: 

  • Prompt injection: Practitioners identified it as the #1 risk, but raw incident counts placed it outside of the top 10. OWASP attributes this difference to the defense effect, in which teams invest significant time into defending against prompt injection, which means that fewer prompt injection exploits occur. OWASP still ranked it as the #1 risk because of prompt injection’s impact on the overall attack surface.
  • Misinformation: Practitioners ranked misinformation as a lower risk than the incident record would indicate. OWASP moved misinformation up two places from 2025 to 2026 because of the risk that inaccurate model outputs that drive tool calls will result in incorrect actions.

Risks that are specific to autonomous agents, including persistent memory and multiagent compromise, are covered in the OWASP Top 10 for Agentic Applications.

A ranked chart showing the differences between the 2025 and 2026 editions of the OWASP Top 10 for LLM Applications. Prompt Injection and Sensitive Information Disclosure remain at ranks 1 and 2. Excessive Agency climbed 3 spots to rank 3, Unbounded Consumption climbed 4 spots from rank 10 to 6, while Improper Output Handling dropped 5 spots from rank 5 to 10. System Prompt Leakage was renamed to Hidden Context Exposure.
Source: OWASP GenAI Security Project, OWASP Top 10 for LLM Applications 2026. 

How Do LLM Security Risks Differ from Traditional Web Vulnerabilities?

Traditional web vulnerabilities come from security issues in deterministic code, which can be identified and patched. They differ from LLM risks because LLM risks come from model behavior, natural-language-based prompts, and autonomous actions that must be addressed at runtime with continuous security checks.

OWASP created a list of LLM risks that are separate from its traditional Top 10 list of web application risks because LLMs blur the line between system instructions and data. Large language models read both system instructions and data as language, which means that an email or document can act as a system command. 

Classic web application security checks, such as input validation, were designed for structured inputs and therefore can’t consistently flag LLM risks. GenAI security practices use least privilege, output checks, and human-in-the-loop strategies to maintain security.

Dimension

LLMs and GenAI Tools

Web Applications

Core components

Probabilistic, natural language-driven models

Deterministic code

System instructions versus data

Blurred together because models read both as natural language

Kept separate, for example as parameterized queries

Common weaknesses

Model behavior, autonomous actions, context exposure

Access control gaps, injection flaws, authentication issues

Where data leaks

Prompts, outputs, reasoning, tool-call arguments, retrieved documents, logs, embeddings, third-party APIs

Database queries, error messages, server logs, misconfigured storage buckets, third-party integrations, API responses

How to fix

Establish guardrails, monitor behavior, test continuously

Find vulnerabilities, apply patches, verify that the fixes were effective

When AI model outputs interact with traditional web components, LLM risks and web app risks merge. For example, unvalidated LLM outputs can cause cross-site scripting (XSS) or SQL injection. That’s why OWASP identifies improper output handling as an LLM risk.

OWASP Top 10 for LLM Applications (2026)

The August 2026 edition of the OWASP Top 10 for LLM Applications updated the rank-ordered list of top LLM risks from the previous year’s edition. The ten risks are:

LLM01:2026 Prompt Injection

Prompt injection occurs when any input that a large language model receives, such as a user prompt, email, document, web page, image, or video, changes how it behaves in an unexpected way. These inputs can cause the model to ignore its system instructions, expose data, or take other unexpected actions.

Prompt injection attacks happen because large language models can’t reliably distinguish between system instructions and data, which means that any content the model reads can be interpreted as a command. This presents a trust-boundary problem that zero trust architectures can help mitigate: zero trust treats every input as untrusted and places limits on the systems that identities, including non-human identities, can access.

Prompt injection comes in two forms: direct injection and indirect injection.

  • Direct injection comes from a user’s prompt.
  • Indirect injection happens when instructions are hidden in content that the model reads on a user’s behalf, including documents, emails, and web pages.

The 2026 edition of the OWASP report includes cross-modal injection, in which malicious instructions are embedded in audio files or images.

What’s an Example of Prompt Injection?

CVE-2025-32711 was a critical zero-click flaw in Microsoft 365 Copilot that let threat actors start prompt injection attacks. In this example, also known as EchoLeak, threat actors could trigger data exfiltration by sending an email that was read by Copilot’s retrieval pipeline.

What Is the Lethal Trifecta?

Before deploying an agent, it's best practice to look for the “lethal trifecta.” The lethal trifecta refers to an AI agent that can access private data, ingests untrusted content, and can communicate externally. An agent with these three attributes meets the conditions for a high-impact exploit, but removing any one of these attributes reduces the threat significantly.

LLM02:2026 Sensitive Information Disclosure

Sensitive information disclosure happens when GenAI apps or LLMs expose confidential, proprietary, or regulated data, like personal health information or login credentials. Data leakage isn’t limited to the model’s output to a user, however. Leaks can come from the model’s reasoning traces, arguments it passes to tools, documents it retrieves, embeddings, and application logs.

Data leakage is a top concern for enterprises as they adopt AI. For example, the ThreatLabz 2026 AI Security Report found that 410 million data loss prevention (DLP) policy violations were tied to ChatGPT alone.

LLM03:2026 Excessive Agency

Excessive agency occurs when GenAI apps are given broader permissions, functionality, or autonomy than is strictly necessary to complete their tasks. It is caused by one or more of three things:

  • Excessive functionality, in which tools have more functionality than needed to complete a task.
  • Excessive permissions, whereby apps can access more resources than necessary to complete the task.
  • Excessive autonomy, in which human checks aren’t required to execute actions. 

For example, let’s say that an IT helpdesk agent is compromised by malicious instructions hidden in a support ticket. If the agent has permissions to reset passwords and edit access groups, the threat actor could leverage those permissions to grant themselves admin access to the system.

Identity-based access controls help defend against this risk, but ThreatLabz researchers found that:

“Across hundreds of enterprise organizations…the average Frontier-AI-Readiness Score was 37 out of 100. Almost all organizations (91%) had publicly reachable AI infrastructure, and none of them had identity-based access controls in place.”

 

 

Read the ThreatLabz 2026 Frontier AI Readiness Report

LLM04:2026 Supply Chain

AI supply chain risks happen when AI models, datasets, adapters, and third-party plugins or repositories that an LLM uses are tampered with or are untrusted. ThreatLabz researchers identified that the AI supply chain is now a key target for threat actors because weaknesses in model files that are used across many custom enterprise LLMs allow threat actors to gain initial access and move laterally within systems.

Four different supply chain attack patterns show how these risks play out: 

  • Slopsquatting: Coding assistants hallucinate package names. Threat actors register those package names in advance, and developers install that malicious code. 
  • Model namespace reuse: If an AI model author deletes their account, threat actors can re-create the account name and publish malicious models under that name. Any pipeline that uses that name to pull models will then download the malicious model.
  • Unsafe serialization: When a model loads, formats like Python pickle can automatically run code. This is a security concern when those models are unverified.
  • Quantization: AI models can pass evaluation checks but exhibit malicious behavior when compressed. It’s a best practice for security teams to test the exact artifact that they intend to deploy.

LLM05:2026 Data and Model Poisoning

Data and AI model poisoning occurs when threat actors or unsafe processes manipulate the data or model artifacts that an AI system relies on. This attack type can also involve threat actors embedding backdoors, bias, or other weaknesses into AI models. Data and model poisoning can happen whenever data is ingested or reused, including during pipelines for pre-training, fine-tuning, embedding, retrieval-augmented generation (RAG), and transfer learning.

Open source and third-party models are a concern, because threat actors can distribute compromised models with embedded backdoors via public repositories. These backdoors stay dormant until they encounter a pre-determined trigger, and unsafe file formats can run malicious code as soon as the AI model loads. Any organization that uses these models can be exposed to these hidden triggers.

If an enterprise uses poisoned datasets to train their models, or uses a compromised open source model to train a custom AI model, it leaves the organization vulnerable to attack.

LLM06:2026 Unbounded Consumption

Unbounded consumption is when an LLM or GenAI tool allows uncontrolled inference, which can lead to outages, increasing costs, or model theft if a threat actor repeatedly queries a model to train a replica model. For example, if a threat actor creates many long and complex queries and repeatedly prompts a public-facing chatbot with those queries, it slows response times for real customers and increases cloud expenses for the hosting company.

LLM07:2026 Misinformation

Misinformation in AI security happens when an LLM outputs false or misleading information that humans or organizations use to make decisions. Let’s say that a compliance team uses an AI assistant to summarize a new regulation’s requirements. If the assistant hallucinates a regulatory requirement that doesn’t exist in the new regulation, but the compliance team doesn’t catch the issue, then the team can waste time and money building controls to comply with a regulatory requirement that doesn’t exist. 

To address misinformation risks in GenAI apps, organizations should ground AI outputs in up-to-date, authoritative sources before using them to guide decision-making.

LLM08:2026 Hidden Context Exposure

Hidden context exposure in GenAI apps happens when system instructions, rules, and other contextual information that users aren’t meant to see are extracted by users or threat actors. One example of a hidden context exposure incident happened in 2023, when a Stanford student used a natural language query to prompt Microsoft’s Bing Chat to reveal hidden system instructions.

Hidden context exposure is especially concerning because threat actors or users can extract, infer, or reconstruct any data in an AI model’s context.

LLM09:2026 Vector and Embedding Weaknesses

Vector and embedding weaknesses in GenAI apps refer to the security flaws in how these apps store and retrieve data, usually in retrieval-augmented generation (RAG). These flaws can allow regular users or threat actors to access or manipulate that data. 

Let’s say that an organization’s internal AI assistant retrieves data from a shared index that includes HR, legal, and financial documents. If a junior employee’s query unintentionally surfaces an executive compensation memo, the security flaw in that AI assistant could surface that information to the junior employee.

Vector databases introduce risks other than access control, because embeddings can be inverted to reconstruct the original text. If a backup of a vector database is leaked, organizations should treat that as a leak of the underlying documents.

LLM10:2026 Improper Output Handling

Improper output handling happens when an AI application passes LLM outputs to other systems or users without first validating that output. A model’s output is only text until another system acts on it, and improper output handling is the mistake of assuming that the text is safe. 

For example, an organization uses an analytics AI assistant to turn natural language questions into queries that are automatically run. If a user sends the assistant a manipulated prompt that gets the model to write a query that exports customer data, the database will execute that query because no check validates it.

How Can Enterprises Secure LLMs?

Enterprises secure large language models by discovering and inventorying all AI systems in use, classifying data that each system can reach, and enforcing least-privileged access controls for users and agents. Other enterprise AI security best practices include inline prompt and output inspection and continuous testing for suspicious model behavior. 

Zero trust for AI deployments helps organizations get ahead of the OWASP Top 10 Risks for LLMs by applying identity-based controls to both human and non-human identities.

LLM Security Best Practices

  • Inventory AI assets and verify provenance (LLM04, LLM05). Identify every model, agent, third-party plugin, MCP server, and embedded AI feature. Verify signatures and pin models to immutable versions rather than mutable tags.
  • Classify data before connecting any AI systems to it (LLM02, LLM09). Enforce  access controls inside retrieval queries, not after retrieval, and only give models access to the data they need to complete their tasks.
  • Apply least-privileged access and human approval controls (LLM01, LLM03). Ensure that agent permissions are aligned with the user’s permissions, rather than a broader service account’s permissions. Require approvals for high-risk actions such as external data transfers or updates to data.
  • Enable inline prompt and output inspection (LLM01, LLM02, LLM10). Security teams should use trained classifiers in addition to pattern matching because regex misses encoded and cross-lingual outputs. Block or set review requirements on outbound content. Disable auto-rendered Markdown images.
  • Never implicitly trust model outputs (LLM07, LLM10). Use parameterized queries, encode outputs for their destinations, and validate outputs before agents use that information to take action.
  • Keep authorization and secrets outside of AI models (LLM08). Don’t put access-control logic or credentials in system prompts. Assume that any context the model contains can be exposed to unauthorized users. 
  • Establish consumption limits (LLM06). Set maximum usage thresholds with per-identity quotas and spending limits.
  • Employ red-teaming for AI systems (addresses all OWASP Top 10 LLM risks). Test all types of attacks and re-test whenever there’s an AI model or tooling update. Align security practices with NIST AI 600-1 and MITRE ATLAS.

How Zscaler Addresses LLM Risks

Zscaler applies zero trust principles to AI security by extending the principle of “never trust, always verify” to AI models, agents, and the data that those systems touch. 

Zscaler AI Security includes: 

 

To learn more, request a demo.

FAQ

No, the OWASP Top 10 for LLM Applications isn’t a compliance standard. It’s a community-driven risk prioritization guide, not a compliance standard or regulatory framework. It ranks the ten most important LLM security risks and provides best practices to help practitioners defend against them. Organizations looking for AI compliance standards should use frameworks such as the NIST AI RMF.

SQL injection exploits applications that include untrusted data in structured queries, and parameterized queries consistently prevent it. It differs from prompt injection because prompt injection exploits AI models, which can’t consistently distinguish between system instructions, data, and user inputs in natural language. There isn’t an equivalent single remediation step that defenders can take to protect against prompt injection attacks, and a layered approach to defenses is necessary.

No, input filtering isn’t a complete defense against prompt injection attacks. Filters and classifiers can be bypassed. For example, EchoLeak chained several bypasses to exfiltrate data. Organizations need to combine input filtering with least privileged access controls, output and link controls, and adversarial testing to block prompt injection attacks.

The main security risks of sending corporate data to public LLMs include sensitive data leaving the organization, loss of control over how business data is stored or used, shadow AI, and the risk of regulatory noncompliance.

A zero trust architecture helps mitigate the top ten OWASP LLM risks because it embeds LLM security best practices into the organization’s existing security architecture. Organizations with this architecture enforce best practices such as treating all AI inputs as untrusted, limiting what resources human and non-human identities can access, and enabling inline traffic and prompt inspection.

The EchoLeak exploit, or CVE-2025-32711, was a zero-click prompt injection flaw in Microsoft 365 Copilot. With this flaw, a threat actor could use an email to make Copilot leak data from its context without user action. EchoLeak was significant for LLM security because it was an example of how indirect prompt injection could work against an enterprise LLM.

The 2026 edition of the OWASP Top 10 for LLM Applications was published in August 2026 by the OWASP GenAI Security Project. There were changes in the rank of LLM risks across the 2025 and 2026 editions of the list, so make sure to check the 2026 edition for the latest risk prioritizations and best practices.

The 2026 edition of the OWASP Top 10 for LLM Applications moved the Excessive Agency risk from 6th to 3rd place and Unbounded Consumption rose four places. Improper Output Handling fell five places. System Prompt Leakage was renamed to Hidden Context Exposure, and Prompt Injection was updated to include cross-modal attacks, which involve images, audio, and video.

Direct prompt injection happens in large language models when a threat actor prompts a model with instructions that override the model’s intended behavior. Indirect prompt injection happens when a threat actor hides malicious instructions in content that an AI model reads while processing a user’s prompt, such as an email, document, or image.

CISOs should use NIST AI 600-1, the Generative AI Profile of the NIST AI Risk Management Framework, for governance guidance, and use OWASP for technical risk prioritization. 

Zscaler recommends that CISOs inventory their AI systems and adopt LLM security best practices such as enforcing least privileged access and human approvals, enabling inline inspection of prompts and outputs, and continuously testing AI systems.