Your AI agents are logging in, opening files, and taking actions every day. Most companies cannot say which accounts those agents use, what they can reach, or how to shut them off. That is an engineering problem with an engineering fix.

Picture a Tuesday morning at a 200-person distributor outside Milwaukee. An AI agent built to answer customer order questions has been running for three months. It works well. Then someone in finance notices it pulled a full customer list with payment terms into a chat transcript. Nobody told it to. Nobody told it not to, either. It simply had access, so it used it.

No hacker was involved. The agent did exactly what its permissions allowed. The problem is that nobody decided what those permissions should be.

That scene is playing out in some form at a majority of companies running AI agents in 2026.

The numbers behind the problem

The Cloud Security Alliance has spent most of this year studying how companies manage AI agent access, and the findings are consistent.

A CSA study commissioned by Zenity, surveying 445 IT and security professionals, found that 53% of organizations have had AI agents exceed their intended permissions. Only 8% said their agents never do. Nearly half, 47%, reported a security incident involving an AI agent in the past year.

A separate CSA study commissioned by Aembit found that 74% of organizations say agents often get more access than they need. 68% cannot clearly tell the difference between actions taken by a human and actions taken by an agent. And 31% have agents running under human user identities, meaning the agent logs in as a person.

A third survey, from identity vendor C1, polled 508 IT and security leaders at U.S. companies with more than 1,000 employees. It found 95% now have AI agents performing tasks autonomously, while only 22% have full visibility into their non-human identities.

These are vendor-sponsored studies, and every sponsor sells a fix. But the pattern across all three is hard to dismiss. Agents are in production. Access controls are not keeping up.

How agents end up with too much access

Almost nobody sets out to give an AI agent the keys to everything. It happens through a series of reasonable shortcuts.

The agent borrows a person’s login. The fastest way to connect an agent to email, a CRM, or a file share is to sign in with the account of whoever is building it. Now the agent has everything that person has, which for an IT lead or an executive can be almost everything.

The service account was built for something else. Teams reuse an existing integration account that was set up years ago with broad rights. Nobody remembers why it can reach payroll.

Connectors make it easy to say yes. Anthropic’s Model Context Protocol, introduced in late 2024 and since adopted by OpenAI, Google, and Microsoft, makes it simple to plug agents into business systems. That is a real benefit. It also means a single setup screen can grant read and write access to a dozen systems in an afternoon.

Nobody owns the agent. CSA found only 15% of organizations say three-quarters or more of their agents have a defined owner. An agent without an owner never gets reviewed, never gets its access trimmed, and never gets retired.

How the major platforms handle permissions

The leading AI vendors take noticeably different approaches, and it pays to know them before you build.

Microsoft 365 Copilot inherits the permissions of the person using it. It can only surface content that user can already open. That is a sound default, but it exposes years of messy SharePoint sharing that was invisible until an AI started reading everything a user technically has access to. Microsoft has also introduced Entra Agent ID to give agents their own identities rather than borrowed ones.

OpenAI’s ChatGPT agent asks for confirmation before consequential actions like purchases or sending messages. That protects against some mistakes, but a confirmation prompt is not an access policy. Users approve prompts quickly when they are busy.

Anthropic’s Claude Code asks permission before editing files or running commands by default, and administrators can set allow and deny rules. Claude’s connectors run on whatever access the connected account has, so the account choice still matters most.

Google’s Gemini Enterprise agents and Salesforce Agentforce both run inside their platforms’ existing permission models, which works well when those models are clean and poorly when they are not.

The common thread: every vendor gives you controls. None of them can decide for you what an agent should be allowed to touch.

Why this matters more in Wisconsin than you might think

Wisconsin’s data breach notification law, Wis. Stat. § 134.98, requires a business to notify affected individuals within a reasonable time and no later than 45 days after learning that personal information was acquired without authorization. When more than 1,000 people are affected, the business must also notify the national consumer reporting agencies.

An over-permissioned agent that exposes customer financial account numbers is not a theoretical AI risk. It is a breach notification problem, with a legal clock attached. And because 68% of companies cannot tell agent activity from human activity, many would struggle to answer the first question an investigator asks: what happened, and who did it?

What good agent permissions look like

The fix is not to stop using agents. It is to treat each agent like a new employee with a very specific job.

Give every agent its own identity. No borrowed human logins. A dedicated identity means every action in the logs belongs to the agent, and you can shut it off without locking out a person.

Grant the least access the task requires. An order-status agent needs to read orders. It does not need payment terms, HR files, or the ability to send email on behalf of the CEO. Start narrow and widen only when a real task demands it.

Separate reading from acting. Reading a record and changing a record are different risks. Many agents need the first and never the second. Where an agent must take action, add an approval step for anything that moves money, deletes data, or contacts a customer.

Use short-lived credentials. Tokens that expire in hours limit the damage from any single mistake. Long-lived keys sitting in a configuration file are how small problems become big ones.

Assign an owner and a review date. Every agent gets a named person responsible for it and a quarterly review of what it can reach. When the project ends, the agent and its access end with it.

Log everything in a way people can read. An audit trail that shows which agent did what, when, and with whose approval is the difference between a quick answer and a 45-day scramble.

Start with an inventory

Most companies do not need a new platform to begin. They need a list. Which agents are running, which accounts do they use, what systems can they reach, and who owns each one? CSA found 54% of organizations have between 1 and 100 unsanctioned agents, so expect the list to be longer than anyone thinks.

That inventory usually takes a few days. It almost always turns up at least one agent running on an executive’s login.

Back to Tuesday morning

Go back to that distributor outside Milwaukee. With a dedicated identity, the order agent would have been able to read order status and nothing else. The customer list with payment terms would have been out of reach. The logs would have shown exactly which agent made which request, and the owner named on the inventory would have caught the drift at the next quarterly review.

Same agent. Same useful work. A much smaller blast radius.

That is the real goal of agent permissions. Not to slow AI down, but to make sure the work it does stays inside the lines you meant to draw. Companies that get this right early tend to move faster later, because leadership trusts the next agent enough to approve it.

Where Earthling fits

Earthling Interactive helps mid-market companies in Madison, across Wisconsin, and throughout the continental U.S. build AI agents that do useful work without carrying more access than the job requires. We design agent identities, scope permissions to the task, wire in approval steps where they matter, and set up the logging that lets you answer hard questions fast. When you already have agents running, we can help you inventory them and close the gaps. Reach out and let’s look at what your agents can touch.