
LLM03:2026
When Tool Permissions Exceed the Task
When Tool Permissions Exceed the Task
An agent can act on its mistakes
In the OWASP Top 10 for LLM Applications 2026, Excessive Agency is LLM03:2026, up from LLM06 in the 2025 list. Ask three questions about internal chatbots and agents this week. Can they complete the task without this tool? Do its credentials grant more access than the person has? Do payments, deletions, and email sends wait for human approval?
Key checks
What to review this week
- Check whether the agent can finish the task without the tool. If it only needs to read, remove delete, send, and deploy permissions.
- Check that actions allowed by the credential fall within the permissions of the responsible person. Shared admin keys and writable production accounts can grant broader access.
- Check that the agent waits for human approval before moving money, deleting data, or sending email. A server’s read-only label alone cannot authorize skipping approval.
- When an out-of-scope tool call triggers an alert, group it with related alerts into one incident. Record the account and tool involved next to the decision. Require approval from the person responsible before isolating or blocking.
Genox groups these alerts into one incident and records the evidence behind the decision. Apply the same three questions to tools used by internal chatbots and agents, and require human approval for isolation and blocking.
References
Reference
- OWASP Top 10 for LLM Applications 2026, OWASP Gen AI Security Project, August 3, 2026. LLM03:2026 Excessive Agency
- Model Context Protocol specification · 2026-07-28, trust tool annotations only from trusted servers. Obtain user consent before calling a tool.
Checked on November 10, 2026