Sofia is invented for this example. She wants something to read supplier email and leave her a summary. A tool that can also hit send is a different animal, and it is wearing the first one’s name tag.
Based on documentation, not a hands-on test.
A map of jobs is Five jobs a rule already does. This article is about what OWASP LLM06:2025 says to limit before a tool can act. Accessed 8 November 2026.
The page’s definition
Agency, on that page, means the system can call a tool and take an action. It is not a compliment. It is a description of what the software is allowed to touch.
The page says an LLM-based system “is often granted a degree of agency by its developer — the ability to call functions or interface with other systems via extensions (sometimes referred to as tools, skills or plugins by different vendors) to undertake actions in response to a prompt.”
It says the decision of which extension to invoke “may also be delegated to an LLM ‘agent’ to dynamically determine based on input prompt or LLM output,” and that agent-based systems “will typically make repeated calls to an LLM using output from previous invocations.”
Excessive agency, in its words, “enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM.” It names three roots: excessive functionality, excessive permissions, and excessive autonomy. It says this differs from insecure output handling, which is about insufficient scrutiny of outputs.
What the page says to remove
Sofia’s summary tool should be able to read. The page’s own mailbox example says it should not also be able to send or delete, unless she has decided that on purpose.
“Limit the extensions that LLM agents are allowed to call to only the minimum necessary.”
“An extension that accesses a user’s mailbox to summarise emails may only require the ability to read emails, so the extension should not contain other functionality such as deleting or sending messages.”
“Utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken.” The example is an app that creates and posts social media content and “should include a user approval routine within the extension that implements the ‘post’ operation.”
The attack scenario on the same page is a mailbox assistant whose plugin can read and also send. A malicious email tricks the model into forwarding sensitive mail. The page says this could be avoided by an extension that only reads, by a read-only permission, or by requiring the user “to manually review and hit ‘send’ on every mail drafted by the LLM extension.”
A fixed email workflow, as Mailchimp’s classic-automation page defines it, is a trigger, a delay, and a message you set. That is not the tool-choosing loop OWASP describes. The Mailchimp page was accessed the same day.
| What it is | Who picks the next step | |
|---|---|---|
| Mailchimp’s object | A trigger, a delay, and a message you set | You did, before anyone received it |
| OWASP’s loop | A system that may call tools, often more than once | The model may decide which tool |
One is a railway. The other is allowed to pick the next station.
Read, draft, and act
- Read the message
- Leave a draft
- Send only if that tool exists
Read. The extension returns the message text. OWASP’s mitigation is that it not also send or delete.
Draft. The text sits for a person. The page’s “hit send” step has not happened.
Act. The send or post runs. The page’s mitigation for that step is approval before the action, or not offering the function.
If one function both reads supplier mail and sends a reply, the page’s mailbox example says the send function is the excess. A prompt line that says not to send leaves the send function in place. Asking nicely is not the same as taking the button off the tool. The page’s stated fix is to remove that function, narrow the permission, or require review before send.
Personal data the tool can see
A summary does not need the whole archive of unrelated people. The regulation’s word for that limit is necessity, not a vibe.
GDPR Article 5(1)(c), consolidation of 4 May 2016, accessed 8 November 2026, says personal data shall be adequate, relevant, and limited to what is necessary for the purposes.
Article 4 defines controller and processor. Article 28 requires a binding arrangement that states the subject matter, duration, nature, purpose, and categories of data when a processor acts for a controller.
Article 22(1) applies to a solely automated decision that produces legal effects or similarly significantly affects a person. Summarising a thread is not, under that sentence alone, such a decision.
Sources
- OWASP LLM06:2025 Excessive Agency, accessed 8 November 2026.
- Mailchimp, About Classic Automations, accessed 8 November 2026.
- Consolidated Regulation (EU) 2016/679, 4 May 2016, Articles 4, 5, 22, and 28, accessed 8 November 2026.