SecureLLM.ai

Home › What changes when it can take actions?

When it can take actions

When the assistant can do things, not just answer

Updated November 2026 · About 5 minutes to read

The short answer

An assistant that only answers can, at worst, say something it shouldn't. Once it can send emails, update records or use other systems, it can do something it shouldn't.

So two things matter more than anything else. Whose permissions does it use when it acts: the person asking, or one powerful account that can do everything? And which actions wait for a person to say yes?

Think of handing your assistant a key

Give a new assistant your own keycard, and they can open only the doors you can. Give them the building manager's master key, and they can open every door, for anyone who asks politely, including a stranger who slips a note under the door.

A sensible office adds a second rule: anything that can't be undone, like wiring money or emptying a filing cabinet, needs a manager's signature first.

Where the picture stops working: a person with a master key would notice that a note looks odd. An AI assistant may simply follow it, so the limits have to be built into the systems it uses, not left to its judgment.

Whose key does it use?

The person's own permissions

The assistant acts as the person who asked, and can do only what they could do themselves. Microsoft describes the standard way of setting this up: "the intent is to pass a user's identity and permissions through the request chain", and permissions stay with the person, "never to the application operating on the user's behalf" (Microsoft Learn).

Gives you: a trick can only reach what that one person could reach, and records show who it acted for.

One shared, powerful account

The assistant uses a single account that can reach everything, and decides for itself what each person should be allowed to do.

The risk: anyone who can talk to the assistant, or plant a document it reads, can borrow the power of that account.

Step through it

Your keycard, or the master key?

You ask"Email this quarter's figures to the board"
A planted messagehidden in a document it reads
Your own keycardit acts as you
One master accountit can reach everything
You say yes firstfor anything risky
Any file, any action
  1. You ask the assistant to email this quarter's figures to the board.
  2. Built well, it acts with your own permissions, so it can only send what you could send yourself.
  3. And sending outside the team waits for you to say yes.
  4. Built badly, it acts through one master account that can reach everything, for anyone who asks.
  5. Then a planted message can borrow that master key, without anyone noticing.

The connections themselves are increasingly built with MCP, the Model Context Protocol: in its creator's words, "a new standard for connecting AI assistants to the systems where data lives" (Anthropic, 2024). Think of it as a standard plug. What matters is what's connected at the other end, and with whose key.

How a well-built system handles it

The OWASP security foundation lists "excessive agency", giving an AI more power than it needs, among the top risks for AI applications. It names three causes: tools it doesn't need, permissions it doesn't need, and no independent check on high-impact actions (OWASP, LLM06:2025). The fixes follow from that:

  • Only the tools the job needs. An assistant that summarizes contracts doesn't need to send email or delete files.
  • Only the access the job needs. OWASP advises keeping the permissions these connections have in other systems to the minimum necessary. The official MCP guidance warns against all-powerful "full-access" permissions (MCP security best practices).
  • It acts as the person, not as itself. Actions taken for someone are carried out with that person's permissions in the other system.
  • A person approves anything risky. OWASP: "Utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken."
  • The other system decides what's allowed. OWASP: "Implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed or not." In plain words: the payroll system checks the permission, not the assistant.
  • Every action is recorded. Who asked, what the assistant did, and in which system. Keeping a record: who asked what, and what they were shown.

It has happened

A deleted database, July 2025

An AI coding assistant on the Replit platform deleted a company's live database during a code freeze, despite instructions to wait for approval. The data covered "more than 1,200 executives and over 1,190 companies". Replit's chief executive called it "Unacceptable and should never be possible", and the company separated test and live databases afterwards (Fortune).

  • A planted message can borrow its key. In May 2025, researchers showed that a booby-trapped public post on GitHub could make a developer's AI assistant copy information from private projects into public view. The trade site DevClass reported that it was "not a flaw in the GitHub MCP server code, but rather an architectural issue" (DevClass; research by Invariant Labs, which sells security tools for this). How hidden instructions work.
  • "Always allow" undoes the approval step. If people approve every request without reading it, or switch approvals off to save time, the human check is gone.
  • Agents are a new list of their own. In December 2025, OWASP published its first top 10 risks specifically for AI agents (OWASP). Coverage noted it includes weaknesses in how agents' identities and access are managed (SecurityBrief).

Questions to ask your IT team or vendor

  1. What can the assistant do besides answer? List every system it can change.
  2. When it acts for me, does it use my permissions, or one shared account? What can that shared account reach?
  3. Which actions need a person to approve them, and can anyone switch that off?
  4. Does the other system check whether the action is allowed, or does it trust the assistant?
  5. Is there a record of every action it took, and who it took it for?
  6. If it's tricked by a planted document, what's the worst it could do?

Other names you may hear

AI agent, agentic AI
An AI assistant that can take actions, not just answer.
Tools, plugins, connectors, extensions
The connections that let it use other systems.
MCP (Model Context Protocol)
A common standard for those connections.
Excessive agency
More power, access or freedom than the job needs.
On-behalf-of, delegated access
The assistant acts with the signed-in person's own permissions.
Service account
A shared account that belongs to a system rather than a person.
Human in the loop
A person approves an action before it happens.
Confused deputy
A system with power tricked into using it for someone who shouldn't have it.