The OWASP Top 10 for LLM Applications: Security Risks in Generative AI Apps
Prompt injection, sensitive data leaks and AI agents with excessive permissions — these are the main risks of LLM-based applications according to OWASP, and how to mitigate them.

Applications built on large language models (LLMs) bring new types of security risk that are not fully covered by traditional web security lists. The OWASP Top 10 for LLM Applications project (now part of the OWASP GenAI Security Project) maps these risks.
The 2025 edition
- LLM01 Prompt Injection — input (direct, or hidden in documents/web pages) that diverts the model's behaviour.
- LLM02 Sensitive Information Disclosure — the model leaks personal data, secrets or internal information.
- LLM03 Supply Chain — risks from third-party models, datasets and plugins.
- LLM04 Data and Model Poisoning — manipulated training or fine-tuning data.
- LLM05 Improper Output Handling — model output used without validation, for example executed as code or queries.
- LLM06 Excessive Agency — AI agents given more functions, permissions or autonomy than needed.
- LLM07 System Prompt Leakage — system instructions containing sensitive information are exposed.
- LLM08 Vector and Embedding Weaknesses — flaws in RAG pipelines and vector databases.
- LLM09 Misinformation — convincing but wrong answers that are trusted as-is.
- LLM10 Unbounded Consumption — unlimited resource usage that drives up costs or causes service disruption.
Mitigation principles
- Treat model output as untrusted input: validate and sanitise it before displaying or executing it.
- Least privilege for agents: limit tools and permissions, and require human confirmation for high-impact actions (sending emails, changing data, transactions).
- Never put secrets in the system prompt; enforce authorisation in the application layer instead of relying on instructions to the model.
- Access control at the retrieval level: documents added to the context must match the user's permissions.
- Limit rates and quotas per user to prevent abuse and cost spikes.
- Monitor and log interactions to detect attack patterns, while respecting personal data protection.
- Test adversarially (red teaming) before and after launch.
AI application security is an extension of application security in general — the fundamentals are the same, but the attack surface now includes natural language.
Sources & references
This article is a summary and interpretation by the Oemah Solution team of the sources below. Refer to the original sources for full details and context.


