Excessive agency
- Attacks
- Standard term
Excessive agency is giving an AI tool more abilities, permissions, data access, or independence than its job needs, which widens the harm an error or attack can cause.
Granting broad access takes one click, and narrowing it takes thought, so tools often keep more than they use. The extra abilities sit idle until the AI misreads a request or follows instructions planted in something it reads, and the AI reaches for them as readily as for the ones its job needs.
Where you’ll meet it
The AI assistant your company just added to its chat workspace answers questions about the employee handbook. Setup switched on every tool the assistant offers, so it can also post in any channel, read private messages, and invite people from outside the company.
What to do
Ask whoever manages the workspace to open the assistant's app settings and switch off everything except reading the handbook and replying to the person who asked. Then ask the assistant to post a test message in another channel, and check that nothing appears there.
Easy to mix up
Prompt injection comes from an outsider who hides text for the AI to obey. Excessive agency comes from whoever set up the tool and gave it more access than its job needs.
Check yourself
A coding agent fixing a typo on your website signs in with the company's admin account. A comment hidden in the site's code tells it to clear out old customer data. Which change, made beforehand, would have done more to stop the delete?
Go deeperOWASP lists it as a top risk and recommends least privilege
OWASP, an open security project, lists Excessive Agency as LLM06 in its 2025 Top 10 risks for applications built on language models, separate from prompt injection, which it lists as LLM01.
The remedy is least privilege, the long-standing security rule that each account gets only what its job needs. For an agent, that means narrow tools, such as a look-up-this-order action in place of full database access.
OWASP also recommends asking a person to approve risky steps. Approving every step can wear down into clicking yes on reflex, so approvals are best saved for steps that can't be undone.
Sources
- LLM06:2025 Excessive Agency, OWASP GenAI Security Project