Three security chiefs, from Intuit, Smartsheet, and testing company ETS, are describing a problem that most businesses have not caught up to yet: nobody owns the AI agents.
An AI agent is a small program an employee sets up to do a task automatically, like pulling numbers from a database, drafting a report, or answering routine customer questions. To do any of that, the agent needs its own login and its own permissions inside the company's systems, separate from the employee who built it.
That is where things fall apart. Atticus Tysen, CIO and CISO at Intuit, describes the risk plainly: agents spread across the company, then get orphaned, then become an entry point for attackers. The employee who built the agent is not doing anything wrong. They just gave the agent more access than it needed, and nobody is watching it after the fact.
Wally Dalyrymple, CISO at ETS and PSI, points to a specific failure point: when an employee leaves, IT switches off that person's own account, but forgets about the agents that person created. Those agents keep running with the same access to company data, sometimes including HR or health records, long after the person who built them is gone.
This is not a small or rare problem. Outside research backs up what these security chiefs are describing. Non-human identities, meaning logins that belong to software rather than people, now outnumber human logins by well over one hundred to one in many company systems, and that ratio has been climbing fast. Roughly half of organizations admit they cannot say who owns a given AI identity, which is exactly the gap that lets agents get orphaned in the first place.
Governments have started responding. In May 2026, cybersecurity agencies in the United States, United Kingdom, Australia, Canada, and New Zealand jointly published guidance telling companies to tighten access controls before rolling out AI agents further. That kind of joint warning from five governments at once is rare, and it signals regulators expect this problem to get worse before it gets better.
The fix, according to Ravi Soin, CIO and CISO at Smartsheet, is treating every AI agent like an employee: give it a name, an owner, a record of what it can access, and a way to shut it off cleanly when its job is done. Smartsheet built an internal system that lists every agent in the company the way an HR system lists every employee.
For most businesses, the practical takeaway is simple. If your company lets staff build AI agents, tools, or automated workflows, someone needs to keep a running list of every one of them, who owns it, and what it can touch. Treat losing track of an AI agent the same way you would treat losing track of a laptop with company passwords saved on it, because that is functionally what it is.