AI & automationSmall-business IT

How to Give Your AI Assistant Its Own Account, Not Your Login

AI agents started over 40% of GitHub pull requests in September. How a small business gives an AI assistant its own limited account instead of the owner's.

By Cybercave Technologies, LLC · Nov 2, 2026

Wall-mounted fingerprint and keypad access reader with a green light beside a glass office door — A door access reader. Illustration only, not a product from any company named here. Credit: Photo by panumas nikhomkhai via Pexels (Pexels License).

Short answer: Don't let an AI assistant work inside your own login. Give it a separate account or connection with only the access its job needs, keep a person on anything that sends, pays, deletes or publishes, and know how to switch it off. GitHub says AI agents started more than 40% of pull requests on its platform in September 2026.

Key points

More than 40% of pull requests (proposed code changes) on GitHub started with an AI agent's first draft in September 2026, up from 17% in March, GitHub said in its Octoverse report on Oct. 28. That was 110 million in one month.
GitHub calls its count a floor, because it only counts agent use it can detect, and Employed AI notes the figure counts which changes began with an agent, not how much of the finished code an agent wrote.
The OWASP security project says an AI agent without a distinct identity of its own makes true least privilege impossible to enforce, and NIST's cybersecurity center proposed on Oct. 27 that agents have identities separate from the people they work for.
The same idea applies outside software: an AI tool connected to your email, files, calendar or website should get its own limited access, not yours.

What to do this week: open the connected-apps page on your main business account, write down every AI tool that has access, and remove any you don't use or can't explain.

Why should an AI agent get its own account?

Because AI agents now do real work inside business systems, and an agent signed in as you can generally do whatever your login allows. GitHub's 2026 Octoverse report found that by September, an agent wrote the first draft of more than 4 in 10 pull requests on the platform. Use of its most hands-off option, where a developer hands off a task for the agent to finish alone, grew 230% in a year. GitHub described review as "how developers shape what comes back."

That is where a separate account helps. If the agent has its own login, it can propose changes and a person with the main account approves them. GitHub's own Terms of Service allow a "machine account" for automated tasks, set up by a person who stays responsible for what it does, and allow one free machine account in addition to a free personal account.

What can go wrong when an AI assistant uses your login?

It can be tricked or simply get things wrong, and when it does, it acts with whatever access your login has. The OWASP Gen AI Security Project calls this "Excessive Agency" and traces it to three causes: the AI has more functions, more permissions or more freedom to act than its job needs.

One OWASP example scenario is an email assistant meant only to summarize messages. Because its connection can also send mail, a booby-trapped incoming email could trick it into searching the inbox for sensitive information and forwarding it to an attacker. OWASP says a read-only connection, or making the user review and press send on every drafted email, could prevent that.

The OWASP Top 10 for Agentic Applications for 2026 lists "Identity and Privilege Abuse" as its own risk. It says an agent without "a distinct, governed identity of its own" leaves an attribution gap, meaning you can't cleanly tell its actions from yours or limit them to what it should do.

Our example (hypothetical): a Holland bakery owner connects an AI assistant to her own email account to sort catering orders. That assistant could then also read the payroll and bank notices in the same inbox, and send mail as her. A separate catering inbox the assistant can read, with the owner pressing send on every reply, gives it the same job with far less reach.

What do NIST and OWASP recommend?

Both point to the same basic rules. NIST defines least privilege as limiting users, or programs acting for them, to the minimum access needed for their assigned tasks. In an Oct. 27, 2026 blog post, NIST's National Cybersecurity Center of Excellence laid out ten initial design tenets for AI agent identity. Among them:

Its own identity. Agents should have identities and access managed separately from human users, so an agent can't broadly impersonate a person and its actions can be traced.
Least access, for the task. Access is limited to what a specific task needs, with no standing rights when the agent is idle.
Access you can take back. Access should be short-lived and revocable, not a password or key that never expires.
A person for big decisions. High-risk actions go to a person for review, without so many requests that approvals become automatic.
Records the agent can't change. Logs of what the agent did are kept somewhere the agent itself can't edit or delete.

NIST notes that current standards and technology can't yet deliver all of these everywhere. Its tenets are a starting point, not a rule small businesses must follow.

How do I set up an AI assistant with its own access?

Start small: one tool, one job, the least access that does the job. These steps are a starting point for any AI tool that connects to a business account.

List what's connected. Check your main business account's linked or connected apps page. Google's Google Account help page shows where to review apps with access to your account, including an "Agent access" filter for apps linked to Gemini or Google's agents, and how to remove access.
Give the AI its own account or inbox. Where the service allows it, create a separate user, mailbox or machine account for the assistant instead of connecting it to the owner's login. Check the service's terms first, as GitHub's machine-account rules show, and note that a paid plan may charge for the extra seat.
Start read-only. When the tool asks for permissions, accept the narrowest choice. OWASP's advice is that a tool that only summarizes email shouldn't be able to send or delete it.
Keep a person on send, pay, delete and publish. Let the assistant draft, then have someone approve. OWASP recommends human approval for high-impact actions, and its example of a tool that posts to social media builds that approval into the post step.
Keep the passwords away from it. Don't paste your own password or a master key into an AI tool. If it needs a key, use one made just for it that you can cancel on its own.
Check its activity and know the off switch. Look over what the assistant did once a week, and write down how to remove its access in one step if something looks wrong.

What should you tell your IT provider?

Ask for four things in writing: a list of every AI tool connected to your business accounts and what each can read, send or change; a separate account or connection for each AI tool, set to the least access it needs; which actions still need a person's approval; and the steps to cut off any one tool's access within the hour.

What should you do this week?

Open the connected-apps page on your main business account, list every AI tool with access, and remove any you don't use. For the ones you keep, ask whether each one needs your login or could work from its own limited account.

For more, see our guides to trying an AI assistant carefully, putting a spending cap on every AI key, and checking who can see what you type into AI.

FAQ

What is an AI agent? It's an AI tool that takes actions, such as sending email, editing files or changing a website, instead of only answering questions in a chat window.

Is it against the rules to make a second account for an AI tool? It depends on the service. GitHub allows machine accounts run by a responsible person, but every service has its own terms, so check before you create one.

Does a separate account make an AI assistant safe? No. It limits what the assistant can reach and makes its actions easier to trace and shut off. It doesn't stop the assistant from making mistakes, which is why a person still approves anything important.

What does "pull request" mean? On GitHub, a pull request is a proposed change to a project's code that someone reviews before it's merged in.

Disclosure: No company or organization named in this article paid for it, reviewed it or saw it before publication. Cybercave Technologies uses AI assistants in its own work and offers AI automation services. Figures and guidance are as published by each source and were checked on Nov. 2, 2026. Not sponsored.