OWASP GenAI Security Project · 2026 (published August 2026)

OWASP Top 10 for LLM Applications: real-world incidents for each item

The 2026 edition is the first to weight its ranking with reported incident data. CollectSEA shows the underlying incidents and turns them into design-time safeguards.

Official OWASP LLM Top 10
  1. LLM01

    Prompt Injection

    Crafted input, direct or hidden in content the model reads, alters the model's behavior.

  2. LLM02

    Sensitive Information Disclosure

    The application exposes personal data, credentials or confidential information through model output.

  3. LLM03

    Excessive Agency

    The model is given more functionality, permissions or autonomy than the task needs.

    and 5 more in the linked patterns.

  4. LLM04

    Supply Chain

    Compromised third-party models, datasets, plugins, extensions or tooling.

  5. LLM05

    Data and Model Poisoning

    Training, fine-tuning or embedding data is manipulated to introduce vulnerabilities or bias.

    No documented incident in the corpus yet.

  6. LLM06

    Unbounded Consumption

    Uncontrolled inference or tool loops exhaust resources or budgets.

    and 2 more in the linked patterns.

  7. LLM07

    Misinformation

    Confident but false output is relied on by users or downstream systems.

  8. LLM08

    Hidden Context Exposure

    System instructions, tool definitions or retrieved content that users should not see are revealed.

    No documented incident in the corpus yet.

  9. LLM09

    Vector and Embedding Weaknesses

    Weaknesses in how embeddings are generated, stored or retrieved.

    No documented incident in the corpus yet.

  10. LLM10

    Improper Output Handling

    Model output is passed to other components (shells, databases, browsers, tools) without validation.

    and 2 more in the linked patterns.

Mappings are CollectSEA's own analysis and are not produced or endorsed by OWASP GenAI Security Project. Item names and summaries paraphrase the official text; consult the source for the authoritative wording.