Excessive Agency
The model is granted more autonomy, permission, or functionality than the use case needs — so it can take consequential actions, especially when manipulated.
What it means
Excessive agency is the risk of giving a model too much power: too many tools, too broad a permission set, or the ability to act without oversight. The harm is not that the model is wrong once, but that a wrong or manipulated decision can delete, pay, send, or change something real.
It is the counterpart to prompt injection: injection is how an attacker steers the model, and excessive agency is how much damage that steering can do.
Why it matters
It rose to #3 in 2026 as agentic applications became mainstream. The more an assistant is allowed to do autonomously, the larger the consequence when it is compromised or simply mistaken.
How we approach it
Minimal, explicit tool surface: the assistant is offered only the endpoints you declare, and the ability to make changes is off unless you enable it for that integration.
It acts as the signed-in user, with that person’s existing permissions — never elevated, never a shared admin credential — so it can never do more than the user could already do themselves.
Consequential actions require explicit confirmation before they run.
A bounded number of tool steps per turn prevents an assistant from spiraling into a long, unsupervised chain of actions.
Frequently asked
Can the assistant take actions on its own?
It can only call the endpoints you declare, and only with write access when you enable it. Anything that changes data is held for explicit user confirmation first.
What permissions does the assistant run with?
The permissions of the signed-in user. It runs within their identity, so it cannot reach data or actions that person is not already entitled to.
How do I limit what the assistant can do?
You control the list of endpoints it is offered and whether write access is enabled. Narrowing that list is the most direct way to reduce its agency.