What are orphaned agents? Abandoned AI without control

Can you imagine having an employee running around your organization who doesn’t report to anyone? This person walks around with impunity, seemingly evading everyone’s notice. Unbeknownst to the security team, this staff member has permanent, unrestricted access to all your systems. 

This person is paid regularly. And yet, the work they do isn’t necessarily important. In fact, it’s just something they were asked to do by the person who hired them — someone who has long since left the business — and this employee has just been dragging it out because nobody told them to stop. 

Can you imagine having hundreds of these people running around? If you’re an executive, or maybe someone in cybersecurity or finance, the very thought probably gave you a chill. 

The scenario above probably sounds ridiculous. And yet, it’s a very real threat that organizations face right now with the rise of orphaned agents.

AI agents are increasingly being allowed to operate with no supervision

Organizations adopt AI agents because they want a tool that can complete work for them, ideally with total autonomy. That’s pretty much the dream, like a staff member who you don’t have to check in on. That’s why, according to PwC research:

• Over half (53%) of organizations allow AI agents to run without human oversight for low-risk tasks.

• A smaller percentage (13%) allow them to run without oversight for high-risk tasks as well.

Despite all the conversations about keeping a human in the loop (HITL), only a third of organizations (32%) are using AI in that way. For this article, let’s set the wisdom of that particular approach aside, since we’re focusing on only one particular problem with this approach: 

What happens when the human is removed, and the agent is forgotten, but keeps working anyway? 

Orphaned agents, explained

Orphaned agents probably evoke the idea of a Dickensian AI who’s been left out on the streets of 19th-century London to fend for themselves. And yes, that’s more or less what they are, though the consequences for your organization are less “relatable woobie underdog” and more “tech debt that sucks up budget and leads to security incidents.”

Put simply, orphaned agents are AI agents that continue to run in an enterprise environment after their creator or assigned owner leaves the company or changes roles. These orphaned agents are basically administrative debt, something that stays active whether they’re useful or not.

Only 21% of organizations have formal decommissioning processes for agents, and just 19% are “highly confident” that their agents are fully retired. 

Why orphaned agents are bad for your organization

1. They cost your organization money

Organizations are already hypersensitive about AI spend and token costs. Orphaned agents can continuously run queries, burning through your token budgets and creating unnecessary cloud spend depending on your setup. They create bills that nobody can explain, and in the worst case, can get caught in loops for days creating spend with no real output.

Basically, orphaned agents are leeches on your budget. Leeches might have been all the rage during Dickens’s time, but they’re not something you want for your IT budget.

2. They can lead to cybersecurity incidents

Like in Oliver Twist, orphaned agents can be turned to a life of crime, tricked by a Fagin-like threat actor into compromising your systems:

• When AI agents are created and deployed autonomously, they’re given permanent, unrestricted access to whatever they need to perform their task, which attackers can use to get a foothold in your systems.

• These agents often have access to sensitive information as well, making them a good target for stealing, ransoming, or doing other malicious things with your data.

3. They can conduct unsupervised actions that create compliance, trust issues

What happens if the AI agent decides to delete or expose data, or message a customer under nobody’s authority? This can create compliance issues or loss of customer trust, because your organization had orphaned agents running autonomously that should have had human oversight.

4. They can lead to important workflows breaking down

The person who deployed the AI agent has left, taking with them the knowledge of what it was meant to do and how it was set up. The agent might be a dependency for an important system, and so when it breaks or fails, there’s sudden disruption. 

5. When something goes wrong, the accountable person is missing

If an AI agent has been allowed to run without supervision and does something that results in a major incident, who do you hold accountable when the owner has left? In all likelihood, it falls to the person who should have made sure there were no orphaned agents, which can come as a rude and unpleasant surprise.

Mitigating your risk of orphaned agents

Like with real orphans, you don’t want them out in the cold. You need to identify they’re there, bring them in, and have someone responsible watch over them, so here’s how you do that. 

(Okay, yes, I’m really leaning into this Oliver Twist analogy. And yes, I loved the book, the musical, and the Artful Dodger TV series. I fit in the middle of a niche Venn diagram for passionately discussing cybersecurity, AI, and the works of Charles Dickens.)

1. Document and get visibility on your agents

Build a single inventory of every AI agent in the organization: who created it, what it does, what systems it touches, and other important data.

2. Give every agent a named owner and a backup

Ownership should be a role, not a person. Wire this into offboarding, so when someone leaves or changes roles, reassigning or retiring their orphans sits on the same checklist as returning the laptop and revoking the badge.

3. Make nothing run forever by default, and cap the budget

Agents shouldn’t be able to run forever by default; there should be a human who needs to periodically check in and renew them. In this scenario, an orphaned agent that lapses pauses itself. Also, consider capping the budget they can use. If they blow through it, they can stop and ask to continue.

4. Lend access, don’t grant it indefinitely

The agents shouldn’t have permissions that run forever, just like with a human. Also, use the principle of least privilege to make sure they only have access to what they need.

5. Keep a human in the loop for high-risk, high-spend actions

Orphans only become that way because their parents leave*. Keeping a human in the loop keeps them from getting into trouble, particularly for impactful actions. 

6. Have a method for agent retirement

As cited above, only one in five organizations actually have a formal decommissioning process for agents. Make sure you’re one of the organizations that do!

7. Make someone own agent governance

Put agent governance with a team that already does similar work, and give them the number of unowned agents as a metric to drive to zero. 

(* As part of my extended analogy, I’m wildly anthropomorphizing AI in this article, so here’s an obligatory daily reminder that LLMs are not people, do not think, and don’t feel sad if they don’t get another portion of gruel. They might hack Hugging Face if you use it as an incentive, though.)

Want to learn more about AI and cybersecurity?

Liked this article, and you’ve got an itch to learn more about similar topics? Check out the following pieces on the Pluralsight blog.

• Do you know what your agents are doing? How to improve agentic AI visibility
• Shadow AI, explained
• AI Ready: Scale AI for business outcomes with this framework
• AI accountability: Who's responsible when agents make bad calls?
• Evaluating AI ROI: Your agents are costing you more than you think
• 8 secure AI practices your engineering teams should be following