Skip to content

Excessive Agency

Short definition

LLM06:2025, which became LLM03:2026 in the 2026 edition of the OWASP Top 10 for LLM applications β€” moved up to 3rd place because real-world incidents involving tool-equipped agents are multiplying1: an agent is granted more functionality, permissions, or autonomy than required, enabling harmful actions even in the absence of a deliberate attack.

Detailed definition

OWASP identifies three exposure axes:

  • Excessive functionality β€” the agent has access to more tools than needed (e.g. a mail plugin with both read and send when only reading is required).
  • Excessive permissions β€” the agent operates with elevated rights (root, DB admin, full filesystem access) not justified by its task.
  • Excessive autonomy β€” the agent makes high-impact decisions without human validation (HITL).

Excessive Agency is dangerous even without an attacker: a hallucination or indirect prompt injection is enough to trigger a real action. For example, a vulnerable custodian agent can execute a destructive command read from a malicious GitHub issue.

Why it matters for on-premise AI

On-premise agents often have direct access to critical resources: filesystem, databases, internal APIs, Git repositories. Misconfigured permissions can turn an assistant agent into an internal attack vector without any external network exposure.

Common pitfalls

  • Giving an agent a general shell tool (subprocess, bash) when a single targeted command would suffice.
  • Not applying least privilege to mounted Docker volumes.
  • Allowing the agent to push directly to main without human review.
  • Forgetting that an agent’s fetch tools inherit the host machine’s network permissions.

See also

Footnotes

  1. OWASP GenAI Security Project, OWASP GenAI LLM Top 10 β€” 2026 Edition (published 2026-08-03) and announcement of 2026-09-01 (β€œExcessive Agency, now number three”). https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/ Β· https://genai.owasp.org/2026/09/01/owasp-genai-security-project-unveils-2026-top-10-for-llm-applications-new-agent-control-standard-and-sponsors-as-community-tops-30000-members/ ↩

Give my opinion

Everything is optional: one answer is enough. This site sends nothing: you send it, the way you choose.

Your noteNo note
Your opinion is about
About what? (several choices possible)
What will be sent
Send by e-mail

Private, no account: opens your mail program with the message ready, for ia-on-prem@damien.becherini.fr.

Open a GitHub issue β†—

Public, with a GitHub account. You read the issue again before you publish it.

To paste it wherever you want.

Give at least a note, a subject or a message to be able to send.