Every Agent Needs Its Own Key: When AI Gets an Identity

27/10/2026

The Spare Key
Dana was leaving for two weeks, so she gave her spare key to Eli, the neighbour’s teenage son. He was responsible and polite, and he’d watered her plants before. ​“Feed the cat, water the ferns, that’s it,” she said.

Halfway through, Eli got invited on a weekend trip. Not wanting to let Dana down, he copied the key at the hardware store and gave it to his friend Jordan to cover for him.
On Saturday, a man in an orange vest knocked on Dana’s door. He said the landlord had sent him to check the water heater. Jordan didn’t want to cause trouble, and the vest looked official, so he let him in.

When Dana came home, her grandmother’s ring was gone.

The police officer asked a simple question: ​“Who was in your home?”

Dana didn’t know. Eli had a key. Jordan had a key. A stranger had walked in through the front door. The neighbor across the hall had seen people come and go, but all she could say was, ​“They had a key, so I figured it was fine.”

Everyone who had come in had come in as Dana.

The next time Dana traveled, she did things differently. The pet sitter got her own key, not a copy of Dana’s, and it opened the front door but not the bedroom. She had a written list of exactly what she was there to do. Nobody else could be let in without a call to Dana first. And the building’s front desk signed in every visitor by name.

Nothing went missing. And if something had, Dana would have known exactly who to ask.

The lock was never the problem. The problem was that everyone who walked through the door looked like Dana.
 

Same Keys, Different Doors.
Dana’s story is the story of many security breaches. Attackers rarely smash down the front door anymore. They walk in with a key, usually someone else’s. Stolen or misused credentials are consistently reported as one of the most common ways attackers get into organizations ‑industry reports like Verizon’s annual Data Breach Investigations Report make this point year after year. Once an attacker is inside with a valid login, they look like a legitimate user, just as everyone who walked into Dana’s home ​‘looked like’ Dana.

Each moment in the story has a direct parallel in cybersecurity, but the ending is where the real damage shows. When the police asked ​“Who was in your home?” Dana had no answer, and many organizations face the same silence after a breach. If everyone shares accounts and the logs only say ​“the key was used,” investigators can’t separate the employee from the attacker. That’s why breaches often go undetected for months, and why companies sometimes must shut down entire systems: they can’t isolate what they can’t identify.

Dana’s fix matches what security teams call identity and access management: 
• Each person gets their own key (unique identities, no shared accounts). 
• The key opens only the rooms they need (least privilege). 
• Nobody gets in on an orange vest alone (verification and multi-factor authentication). 
• The front desk signs everyone in by name (audit logs). 

None of this makes Dana’s home impossible to break into, but it makes misuse much harder and much easier to trace.

That brings us to AI agents. We are now handing out spare keys to software that acts on its own, works around the clock, delegates tasks to other agents, and can be fooled by a convincing orange vest hidden inside a document. If we give agents borrowed keys and no sign-in sheet, we’ll recreate Dana’s first trip at machine speed and scale. The lesson from the house applies even more strongly to agents: every actor needs its own key, limited access, verified introductions, and a record of who did what.
 

From Visitors to Residents: Why AI Now Needs an Identity
For most of its short history, AI in the workplace has been a visitor. A chatbot answers a question, drafts a paragraph, or summarizes a report, and then it’s done. It never holds a key. It can’t open a door, move a dollar, or send a message in anyone’s name. A visitor like that doesn’t need much of an identity, because it can’t do much harm on its own. Having said that, any assistant that accesses private data or services still needs clear authentication and authorization.

That has changed. Agentic AI doesn’t just answer; it acts. It books meetings, queries databases, sends emails, moves money, writes and deploys code, and hands work to other software and other agents. To do any of that, it needs what Eli needed: a key. And the moment an AI holds a key, it stops being a visitor and becomes a resident of your systems. Every organization now must ask about its agents the same questions Dana should have asked about her house: Who is this? Who let them in? Which rooms can they open? And who answers for it when something goes missing?

The challenge is that our locks and sign-in sheets were built for people, not agents. Traditional identity management assumes a human logs in, or a piece of software runs the same way every time with the same fixed permissions. Agents break those assumptions, and each break is a reason identity management must evolve.

First, an agent typically acts on someone’s behalf, so there are two identities at the door: the agent and the person who sent it. Like Eli, the agent carries someone else’s authority, and if we can’t tell the two apart, every action looks like it came from the human. An agent also shouldn’t quietly end up with more power than the person who delegated the task.

Second, agents come and go far faster than people do. An agent might be created for a single task and disappear minutes later, sometimes thousands of agents at a time. Our systems for handing out and taking back keys were designed around hiring and offboarding employees, not around a stream of short-lived workers.

Third, agents make their own decisions. A traditional program follows the same steps every time, so giving it fixed access is predictable. An agent decides what to do based on its goals and often untrusted content, using any tools or actions it has access to, which makes a master key far more dangerous than a key to one room. 

Fourth, agents hand off to each other, just as Eli handed the key to Jordan. Agent A asks Agent B, which calls a tool, which calls another system. Each handoff blurs who authorized what, and by the end of the chain, nobody can say whose authority the final action carried.

Finally, agents can be fooled. Malicious instructions hidden inside an email, a web page, or a document can trick an agent into misusing access it legitimately holds. This is the orange vest problem, which security professionals call the ​“confused deputy”-caused by indirect prompt injection: a trusted insider manipulated into doing an outsider’s bidding. The more keys an agent holds, the more damage a convincing vest can do.

Put together, these changes mean we can’t keep treating agents like visitors or like ordinary software. They need what Dana’s pet sitter eventually got: their own key, access to only the rooms the job requires, verification before anyone new gets let in, and a sign-in sheet that records every entry by name. That is identity management for AI, and it is quickly becoming the foundation for trusting agents with real work.
 

Good Practices for Giving AI the Keys
The foundation is a unique identity of its own. Every agent, or at least every type of agent, should be registered as a distinct identity rather than borrowing a person’s login or sharing a generic service account. That’s what makes it possible to tell who did what later.

For proving who it is, a strong option is a cryptographic workload identity, such as a certificate or a signed identity token issued automatically by the platform the agent runs on. Frameworks like SPIFFE (Secure Production Identity Framework for Everyone), and the managed identities built into AWS, Azure, and Google Cloud, do this. These are the ID badges with holograms: hard to steal, hard to fake, and good for agents verifying each other. These frameworks also promote automated issuance, short lifetime, secure storage, and verification of identity.

For acting on a person’s behalf, the agent should use delegated, scoped tokens, typically issued through OAuth or equivalent. A good token records both the agent and the human who authorized it. It also covers only specific actions (“read calendar,” ​“approve invoices under $10,000”) and expires quickly, often within minutes or hours. 

Access should be just-in-time. Instead of standing access to everything, the agent requests access when a task starts, and the access disappears when the task ends. This matters especially for agents that exist for only a few minutes.

For high-risk actions, such as moving money, deleting data, changing bank details, or deploying code, the agent should hold a key that requires human approval.
 

Every Agent Needs Its Own Key
Dana’s ring didn’t go missing because her lock was weak. It went missing because she couldn’t tell who had walked through her door. Her key had been copied, passed along, and used by a stranger in an orange vest, and every one of them presented like Dana. The fix wasn’t a stronger lock. It was knowing exactly who held each key, what it opened, and who let them in.

AI is now standing at that same door. A chatbot that only answers questions doesn’t need much of an identity. An agentic AI is different. It takes actions: it books meetings, queries databases, sends emails, moves money, writes and deploys code, and calls other software or other agents. To do those things, it needs credentials and permissions. At that point the agent becomes an actor in your systems, and every organization has to answer questions it used to ask only about employees and service accounts. Who is this? Who authorized it? What is it allowed to touch? Who is accountable when it does something wrong?

If we can’t answer those questions, we’re back on Dana’s first trip, except the house is our entire organization, the visitors work around the clock, and they can hand the key to a thousand others in seconds. If we can answer them, agents become what Dana’s pet sitter was on the second trip: genuinely helpful, trusted with real work, and accountable for what they do.

The goal isn’t to keep agents out. It’s to make sure that when they walk through the door, they walk in as themselves, with their own key, opening only the rooms they need, and signed in by name. When AI gets an identity, we will only then know who did what.