Internet Security Auditors Blog

OWASP Top 10 for LLMs: The 10 risks that turn an AI-powered application into a new attack surface.

Written by Internet Security Auditors | Apr 9, 2026 9:05:35 AM
For a while, the conversation around generative AI security revolved almost entirely around jailbreaks. However, the real problem begins a bit later—when the model doesn’t just respond, but also queries documents, invokes tools, calls APIs, accesses corporate systems, or performs actions on behalf of the user. At that point, the LLM stops being a simple conversational engine and becomes a new attack surface.

That’s where the OWASP Top 10 for Large Language Model Applications becomes especially valuable. This guide organizes the most relevant risks that emerge when a model is integrated into real applications, with real data, and real business impact. It doesn’t just talk about malicious prompts—it addresses permissions, supply chain issues, information leaks, insecure RAG, output manipulation, misinformation, and uncontrolled resource consumption.

The 2025 edition is particularly interesting because it reflects the ecosystem’s maturity. OWASP keeps prompt injection as the top risk but gives more weight to problems that are now critical in enterprise deployments: system prompt leakage, weaknesses in vectors and embeddings, excessive agency, and unlimited consumption. Translated into business language: the risk is no longer just in what AI says, but also in what AI sees, can touch, can execute, and can cost.

What is the OWASP Top 10 for LLM Applications?

The project was launched in 2023 as a community-driven initiative to bring visibility to the specific security risks found in applications built on top of LLMs. The 2025 edition, published in March of that year, expands and refines the initial work with new categories and a much more aligned focus on current use cases: corporate assistants, copilots, agents, RAG architectures, and automations connected to third‑party systems.

The usefulness of the Top 10 is twofold. On one hand, it provides technical teams with a clear taxonomy to model threats, prioritize controls, and design security testing. On the other hand, it helps business and risk leaders understand that an AI‑powered application does not introduce a single new problem, but rather a combination of classic and emerging threats that amplify one another.

What changes in the 2025 edition

OWASP highlights several fundamental changes in the 2025 edition. The first is the evolution of the former denial‑of‑service logic into a broader category: unlimited consumption. In LLM environments, poor management of context, tokens, tool calls, or model access doesn’t just affect availability; it can also drive up costs and silently degrade service quality.

The second major change is the explicit inclusion of vector and vector‑representation weaknesses, a direct response to the rise of RAG architectures. When an application’s security depends on how it generates, stores, and retrieves embeddings, the vector store stops being a neutral component and becomes a critical asset.

The third important change is the introduction of system prompt leakage. Many applications assumed these internal instructions were opaque to the user. OWASP makes it clear that this assumption is unsafe. If the system prompt can be revealed, it can also expose internal logic, credentials, security constraints, and operational details that enable new attack paths.

Finally, the excessive agency category gains weight due to the growing use of agents and plugins. When the model has autonomy to act on other systems, any control failure stops being a simple response error and becomes an unsafe action with operational impact.

The Top 10 at a Glance

ID  Riesgo
LLM01  Prompt Injection
 LLM02   Sensitive Information Disclosure 
 LLM03   Supply Chain Vulnerabilities 
 LLM04   Data and Model Poisoning 
 LLM05   Inadequate Output Handling 
 LLM06   Excessive Agency 
 LLM07   System Prompt Leakage 
 LLM08   Vector and Embedding Weaknesses 
 LLM09   Misinformation 
 LLM10   Unlimited Consumption 


1. Prompt Injection
Prompt injection remains in the top position for a simple reason: external content is not just context — it can also be an attack vector. An email, a webpage, a PDF, a ticket, or a document inside a RAG system may contain hidden instructions capable of altering the model’s behavior. The threat no longer lies solely in what a user types directly, but in any source the model processes and interprets.

 Objectives 
➡️ Treat all user input as untrusted.
➡️ Use strict input validation, privilege separation between the model and backend systems, and human review for sensitive operations.
➡️ Do not allow LLM‑generated content to directly trigger privileged actions without verification.
 

2. Sensitive Information Disclosure
Sensitive information disclosure highlights that an AI‑powered application can leak secrets without requiring a traditional intrusion. All it takes is poor training practices, weak orchestration, improper retrieval, or overly broad access. When an LLM connects to repositories, document stores, CRMs, tickets, or source code, the exposure surface expands immediately.

 Objectives 
➡️  Expose only the data that an LLM truly needs.  
➡️ Apply output filtering to detect and mask sensitive patterns before responses reach the user.
➡️ Periodically audit training and retrieval data flows.

 
3. Supply Chain  
In AI environments, the supply chain includes dependencies, datasets, base models, LoRA adapters, third‑party repositories, conversion processes, and the provider’s terms of use.

 Objectives 
➡️  Carefully review all third‑party models, plugins, and data sources.
➡️  Prefer models from reputable providers with documented security practices. 


4. Data and Model Poisoning
Cases such as PoisonGPT, attacks on model‑conversion processes, or the abuse of external components show that the problem can arise at any point in the pipeline.

 Objectives 
➡️  Carefully validate all training data. 
➡️  Use anomaly detection during training. 
➡️  Prefer fine‑tuning over full retraining whenever possible, and monitor model outputs in production.


5. Inadequate Output Handling
In parallel, inadequate output handling serves as a reminder that an LLM’s response should never be blindly chained to another automatic action without deterministic validation.

 Objectives 
➡️  Use parameterized queries for database interactions, sandboxed code‑execution environments, and implement strict output validation.


6. Excessive Agency
OWASP refers to excessive agency as one of the most important issues today. The risk is not just that the model might make a mistake; the real problem is that such a mistake may have permission to create, delete, send, approve, purchase, or modify something.

 Objectives 
➡️  Apply the principle of least privilege rigorously.
➡️  Grant LLM agents only the minimum permissions required for each specific task.  
➡️  Implement human‑approval gates for high‑impact actions. 

 
7. System Prompt Leakage 
System prompts contain information about intellectual property, security configurations, personality definitions, and operational instructions that developers consider confidential. Attackers can often extract partial or complete system‑prompt content.

 Objetivos 
➡️  Use robust access controls, avoid embedding credentials or sensitive business logic in prompts, and monitor extraction attempts.


8. Vector and Embedding Weaknesses
Weaknesses in how embeddings are generated, stored, or retrieved can allow attackers to manipulate retrieval results, inject malicious content into the knowledge base, or exploit similarity‑search algorithms to extract sensitive stored documents.

 Objectives 
➡️  Validate all content before it enters the knowledge base.
➡️  Implement query filtering to prevent unauthorized document retrieval, and monitor vector‑search patterns for unusual or suspicious behavior.

 
9 y 10. Misinformation and Unlimited Resource Consumption
Misinformation and unlimited consumption introduce an important idea: an AI‑powered application can also cause harm when it produces convincing falsehoods or when it consumes resources without control. This affects reputation, operations, decision‑making, and cost. In many corporate environments, an incorrect output or an inefficient architecture can be just as damaging as a traditional technical vulnerability.

 Objectives 
➡️  Base model responses on verified and authorized data sources. 
➡️  Display clear confidence indicators and source citations. 
➡️   Implement human review for critical decisions and educate users about the limitations of AI‑generated content.

 
Apply per‑user limits and maximum output‑token limits. Monitor unusual usage patterns and implement circuit breakers to stop uncontrolled processes.

What organizations should review today

Before deploying an assistant, a copilot, or an agent connected to internal systems, organizations should review at least six aspects. First, clearly separate trusted from untrusted content and label external sources. Second, limit privileges and access following the principle of least privilege. Third, introduce deterministic validations on model outputs before using them in other components. Fourth, require human approval for sensitive actions. Fifth, audit the supply chain of models, datasets, libraries, and providers. And sixth, monitor consumption, costs, context usage, and tool calls to detect abuse or operational drift.

In other words: less fascination with the demo, and more discipline in architecture, governance, and security. The OWASP Top 10 for LLM Applications is not a list to memorize; it is a guide to help you ask the right questions before allowing an AI system to read, decide, or act inside the business.

▪️Separate trusted and untrusted content in prompts and RAG flows.
▪️Apply least privilege to tools, connectors, APIs, and agents.
▪️Validate model output deterministically before reusing it.
▪️Maintain human approval for critical or irreversible actions.
▪️Audit providers, models, datasets, and supply‑chain dependencies.
▪️Control consumption, costs, context usage, quotas, and tool calls.

Conclusion 

The great merit of the OWASP Top 10 for Large Language Model Applications is that it forces organizations to view generative AI for what it already is in many environments: a critical software component. And when a critical software component accesses data, processes, and tools, security cannot be added at the end of the project. It must be part of the design from day one.

Because the problem doesn’t start when the model “hallucinates.” It starts when that hallucination has access to a system, a sensitive piece of data, or a business decision.

References
OWASP GenAI Security Project - LLM Top 10
OWASP Top 10 para Aplicaciones de LLM - Versión 2025 (traducción al español)

 

José Antonio Linio

OSCP
Security Auditor 
Audit Department

 

 Ismael de Frutos

CRTP
Security Auditor 
Audit Department