
LLM01:2026
Prompt Injection,
Prompt Injection,
Instructions in chat and documents
The OWASP Top 10 for LLM Applications 2026 was published August 3, 2026. Prompt injection remains LLM01:2026. Distinguish instructions a person types into chat from instructions embedded in email, wiki pages, and attachments the model reads. Require human review before sending, deleting, or making an external call.
Key checks
What to review this week
- List the inputs internal chatbots and agents read, including chats, emails, wikis, attachments, and search results.
- Check whether the interface distinguishes direct user instructions from instructions embedded in documents. Keep document text in a separate field from user instructions.
- Check that a person can review the target and contents before the agent moves money, deletes data, or sends email.
- Group chatbot and agent alerts into a single incident and record the evidence behind the decision. Require approval from the person responsible before isolating or blocking.
Genox groups related alerts into one incident and records the evidence behind the decision. Apply the same process to internal chatbots and agents by recording calls prompted by document content as incidents and requiring human approval for response actions.
References
Reference
- OWASP Top 10 for LLM Applications 2026, OWASP Gen AI Security Project, August 3, 2026. LLM01:2026 Prompt Injection
Checked on November 10, 2026