Skip to the stories
Microsoft 365 & SharePoint

AI Agent Sprawl: Finding and Governing the Orphans in Your Estate

  • Kelsey Ziser
  • 695 views

Orphaned AI agents slip past IT visibility and inherit access they shouldn't have. Here's how identity, ownership and audit teams can get them back under control.

AI Agent Sprawl: Finding and Governing the Orphans in Your Estate

Traditional identity and access management was built for people and for relatively static machine accounts. AI agents break both assumptions. They appear on laptops, get wired into third-party software, run under service accounts and quietly outlive the people who set them up. CIOs and CISOs are now rewriting their identity playbooks to deal with that, and the work starts with knowing what's actually running.

Why orphans are the real problem

The term that keeps coming up is "AI orphans": agents operating outside IT's line of sight, or agents left behind when someone leaves the company. The second case is the one that trips teams up. When a former employee's user account is disabled, the agents tied to that person don't automatically disappear. IT has to decide whether to shut down every agent linked to that account or keep managing them, and that decision isn't always obvious.

Employees aren't the only source of sprawl. Enterprise teams also have to secure agents that arrive inside third-party vendor software. Some of those agents run against a service account rather than a single human, which muddies permission decisions because there's no one clear person to point to.

Complexity grows once agents have several owners. Hanah-Marie Darley, chief AI officer at AI agent management company Geordie AI, said many larger organisations run multi-use or multi-workflow agents tied to multiple human users. One agent may therefore need several sets of permissions, depending on who is using it and for which workflow.

Darley also pointed to the dynamic nature of agent access as a major shift from older machine identities. "It's really different from machine identities because they used to be just kind of set — they would always access the same resources," she said. The total volume of tool and system access, and how often it changes, is what makes the job hard.

Visibility comes first

Dormant agents are a quiet risk. An employee might switch from one agent to another and never tell IT. The old one keeps its connections and sits unmonitored in the environment. Darley's point is blunt: if you can't see your agents, you can't govern them. That applies to internally built agents and third-party ones alike, which means development teams and the security staff managing vendor tools need to work from the same picture.

Darley said most organisations don't have a unified view of agent operations. A useful inventory answers four questions: where agents exist, whether they're active, how their identities behave in practice, and whether they need to be retired.

Allen Houchins, CIO of Fleet Device Management, framed the scope even more broadly. Visibility, in his view, means knowing how employees use corporate devices, down to the detail. "It's really about visibility and inventory of all of your devices — and not only the agents and the applications that are running but really getting into the finer details of agent instruction files, plugins, MCPs, extensions, even local [AI] models that people are installing," he said. "If you don't have this visibility, you can't govern it, control it or make any business decisions about it."

Once inventory is complete, gaps often show up where AI could add value. Before deploying anything new, though, CIOs should define the business outcome they expect and how they'll measure performance. Otherwise the inventory just becomes a longer list of things nobody can justify.

Tie lifecycle management to outcomes

Proving ROI means managing an agent's entire lifecycle, not just its launch. Darley's framing is that lifecycle work has to be tied to value. "You have to tie all lifecycle management to value because ultimately you judge an agent by outcomes," she said. That assessment needs to be continuous rather than a one-off measurement at go-live.

The lifecycle also has an end. Some agents stop delivering the outcomes they were built for, or become too expensive to run. Good lifecycle management should flag those cases in near real time so financial and operational damage stays small. One practical step is to write retirement policies in advance: decide when an agent gets decommissioned or deleted, close off its permissions, and confirm it isn't still talking to other agents.

Deletion isn't always the answer. Rob Hughes, CISO at IAM solution provider RSA, noted that teams may prefer to archive an agent instead. The reason: "there was all this thought and knowledge that went into it and the next person in that role might have an interest in it." Whichever route you choose, there should be a clear, documented process for updating or removing the agent's access.

Give every agent an owner

Human employees have managers, and agents need someone equivalent. Darley said many of her customers assign an accountable owner, or a human "parent user", to each agent. Houchins put it plainly: "You need to give every single agent a named owner."

Raj Ananthanpillai, CEO of identity screening company Trua, goes further. He recommends giving employees a "trust token", which they obtain after biometric and ID document checks. That links a verified human to any agent acting on their behalf, such as one making an online transaction. He compared the idea to TSA Precheck, where travellers are vetted in advance and receive a unique number that can be validated at booking. "People should bind all of their agents in the company to the individual," Ananthanpillai said.

Scale changes the maths. Some of Darley's larger customers run thousands of agents, so manual review by an AI governance team isn't realistic. She said speeding up management at that scale requires independent AI systems to help do the work.

Control access at the API and MCP layer

Governance has to reach the points where agents actually touch data. Hughes said proper control at the API or Model Context Protocol level is essential for keeping agents inside a controlled environment. He described a case where an agent identity connected to an API, and the API gave it unfettered access to the organisation's data. That exceeded what the agent identity had been granted and what it was expected to have. The intended access was that of the person who created the agent.

Danny Jenkins, CEO of cybersecurity company ThreatLocker, agreed that IT leaders need explicit policies covering which data and systems AI can reach. He urged teams to define permissions clearly at several levels. "Make sure that you are very clearly defining, at a file-permission level, at an application-permission level, a ring-fencing level, at an AI control level — whatever way you want to do it — which files, which documents, which other applications on your computer this AI agent can talk to, and when it does talk to them, that those permissions flow through," he said.

Audit continuously, and look at the portfolio

Audit processes also need to change. Hughes argued that agent reviews can't be a one-off exercise. Treating agents like standard service accounts and reviewing ownership quarterly or yearly no longer works. "There's so much more complexity wrapped up in these new non-human identities than prior," he said. "Really what you need is a continuous review of what the agents are doing and making sure they're not going out of the bounds that you set."

CIOs also need to look at both the micro and macro levels. An agent can perform well in isolation and still be a poor use of money across the whole estate. Darley described a customer that discovered about 19% of its AI budget was going on automating scheduling alone. Those agents worked well individually, but as a fifth of total AI investment they weren't delivering the outcomes the business needed. Assessing agents as a portfolio, not just one at a time, is how that kind of imbalance gets spotted.

The pressure to deploy AI quickly means some agents will slip through regardless. Darley's advice is pragmatic: "Everybody at some point will have orphan agents," she said. "That's OK, so don't feel bad. But think about how you want to manage the lifecycle of agents rather than just getting them up and running."

My read is that the organisations that handle agent sprawl well will be the ones that treat agents as identities with owners, lifecycles and audit trails from day one, rather than as experiments that can be cleaned up later. Start with an inventory, attach a named human to every agent, tie each one to a measurable outcome, and set a retirement rule before the next wave of deployments lands.

Source: Kelsey Ziser · www.informationweek.com

Neighbouring cells