OWASP LLM Top 10

OWASP LLM Top 10

Table of Contents

Quick Answer

The OWASP LLM Top 10 is a beginner-friendly way to understand common security risks in applications that use large language models. It covers issues such as prompt injection, insecure output handling, sensitive data exposure, excessive agency, overreliance, and model theft.

What is the OWASP LLM Top 10?

The OWASP LLM Top 10 is a community-driven list of important security risks for applications that use large language models. It helps developers, security teams, and beginners understand how AI application risks differ from normal web application risks.

Why LLM Application Security is Different

LLM applications often combine prompts, retrieved documents, plugins, APIs, model outputs, and automated actions. This creates new trust boundaries. Traditional controls still matter, but AI applications also need prompt isolation, output validation, least-privilege tools, cost controls, and human review for sensitive actions.

The 10 Risks Explained in Simple Words

OWASP LLM riskSimple meaningBasic defense
Prompt InjectionUntrusted text manipulates model behaviorSeparate instructions from data and validate actions
Insecure Output HandlingModel output is trusted unsafelyValidate and sanitize before use
Training Data PoisoningBad data influences model behaviorCurate, monitor, and validate training data
Model Denial of ServiceCost or resource abuseRate limits, quotas, and monitoring
Supply Chain VulnerabilitiesRisky models, datasets, or componentsVet dependencies and providers
Sensitive Information DisclosureSecrets or private data appear in outputsData minimization and access control
Insecure Plugin DesignTools or plugins expose risky actionsLeast privilege and approval checks
Excessive AgencyAI takes too much action autonomouslyScope permissions and use human approval
OverrelianceUsers trust incorrect AI outputVerification and human review
Model TheftUnauthorized model access or copyingAccess control, monitoring, and abuse detection

Which Risks Matter Most for Developers?

Developers should start with prompt injection, insecure output handling, sensitive information disclosure, plugin/tool permissions, and excessive agency. These risks appear when an LLM application is connected to user data, documents, APIs, code, or business workflows.

Which Risks Matter Most for Users?

Users should be careful about entering sensitive data into AI systems, trusting unverified AI output, and allowing AI tools to perform important actions without review. Human verification remains important even when an AI response sounds confident.

How to Start Securing an LLM Application

  1. Map where prompts, retrieved content, outputs, tools, and data flows exist.
  2. Separate trusted instructions from untrusted content.
  3. Validate output before it is used in HTML, SQL, code, or tool calls.
  4. Apply least privilege to plugins, APIs, and agents.
  5. Use logging, rate limits, monitoring, and human approval for sensitive workflows.

Recommended Learning Path

Start with the AI Security Roadmap, then study Prompt Injection, Cross-Site Scripting, SQL Injection, and Penetration Testing. LLM security is easier to understand when the web security fundamentals are clear.

FAQs

The OWASP LLM Top 10 is a list of important security risks for applications that use large language models, generative AI, plugins, tools, agents, or AI-connected workflows.

No. Developers, security teams, product owners, and learners can all use it to understand common LLM application risks and safer design patterns.

Prompt injection is a good first topic because it explains how untrusted text can influence an LLM application, especially when tools or documents are connected.

LLM security includes normal web security controls but adds prompt isolation, output validation, tool permissions, excessive-agency controls, data minimization, and human review.

Start by mapping AI data flows, identifying tools and permissions, reviewing prompts and retrieved content, validating outputs, limiting sensitive actions, and monitoring AI workflows.

Sources and further reading