Who’s Accountable When an AI Agent Acts? Identity Governance with SailPoint’s Mikael Wikström

Louis Zezeran 19. nov. 2026

In June 2026, Estonia’s Eesti.ai advisory council made a proposal no other country had made: give AI agents an official digital identity, an AI-isikukood, modelled on the personal ID code every Estonian carries. Prime Minister Kristen Michal approved it and called it a world first. Within days, Estonian lawyers were writing op-eds arguing that what AI agents really need is collars, not passports. 

It’s tempting to treat this as a local curiosity from the world’s most digital state. It isn’t. The same argument is playing out inside every enterprise identity team. AI agents are requesting access, calling APIs and acting for employees who may not even know those agents exist. Who are these agents? What can they touch? And when one does something it shouldn’t, who answers for it? 

For this episode of Cybercast by NEVERHACK, recorded live on the floor of the Nordic Baltic Security Summit, Ronnie Jaanhold of NEVERHACK Estonia put those questions to someone who has to answer them in a shipping product, not just an opinion. Mikael Wikström has worked in identity for over three decades, the last ten at SailPoint. He is a Cybersecurity Advisor there and leads the solutions engineering team across the Nordics, the Baltics and Benelux. 

The result is a short, practical conversation that will change how you think about identity in the age of agents. 

The biggest milestone in 30 years of identity 

Mikael started in identity when the field moved from simple account management to treating security in a human identity context. He calls AI the biggest milestone he has seen since. The reason is scale. 

Take a mid-sized enterprise with 10,000 users and 2,000 systems. That’s 200,000 accounts straight away, and once you map what each account can reach, you have millions of access points between people and data. Then add the non-human identities: service processes, API credentials and, increasingly, AI agents. No team of people can analyse that by hand. SailPoint first used machine learning in its platform 13 years ago for exactly this reason, and today AI is what makes it possible to spot risky patterns in all that complexity. 

AI also changes how identity platforms are built. Mikael’s example is workflows: what should happen when someone leaves the company or moves department? Historically these workflows were slow to build and error-prone. Now administrators can describe in plain language what they want, for example “start an access review with the new manager, notify them, and disable the account until it’s complete”, and an agent builds the workflow for them to check and adjust. 

So what actually is an identity? 

When we hear “identity”, most of us picture a person. Mikael’s definition is much wider, and it is the foundation of the whole episode. 

Human identities are the simple part. Non-human identities are where the landscape gets crowded. An AI agent is an identity. So is a process that uses several accounts to reach a database. So, Mikael argues, is a single credential, such as a key to a REST service, because it is an entity that can access data. 

The principle underneath all of this is the line to remember from the episode: you cannot address non-human identities without humans. Every agent, process and credential needs to sit in a human context, with an accountable owner, so that someone is responsible for what it can reach and what it does. 

Does an AI agent need its own ID? 

This is where the episode comes back to Estonia. Mikael’s answer is more nuanced than yes or no. 

Yes, agents should be treated as identities in their own right. No, they don’t necessarily need a new central ID, because they already have one. An agent spun up in Microsoft Entra, AWS or Google Cloud already has a unique identifier in that system. And agents aren’t permanent objects: one might live for milliseconds, another for ten years. SailPoint’s approach is to connect to those platforms, detect every agent as it is created, find the accounts it uses and the access those accounts carry, and then tie all of it to a responsible human who has to verify that it’s correct. 

Should agents always act on a human’s credentials, or independently? Both, Mikael says, for the foreseeable future. Acting on behalf of a user is fine for many tasks. But an agent that monitors incoming invoices around the clock is autonomous, so it needs its own identity and its own access. Either way, the same rule applies: you have to be able to detect it, govern it and react to it. 

Why twice-a-year access reviews no longer work 

Traditional identity governance follows a familiar rhythm. A manager reviews their team’s access two or four times a year and confirms everyone has what they should. 

With AI agents, Mikael says, that is “absolutely not enough”. Agent access is in constant flux. What’s needed is a real-time policy engine that: 

  1. Detects agents the moment they’re created 
  1. Maps what data they can reach, such as files containing personal data, sensitive information or intellectual property 
  1. Applies policies that set what that agent is allowed to access and analyse 
  1. Reacts the moment it steps outside those boundaries, by revoking the credentials or stopping the agent entirely 

Mikael gives a concrete example: a policy says an agent may not include personal data in its responses. The engine watches the output, and when a personal ID number appears, it treats that as proof the agent has too much access and shuts it down. 

The human owner is still essential, because accountability doesn’t go away. But people can’t react in microseconds, and machines can. Humans set the rules and answer for the agent; the platform enforces the rules at machine speed. 

Group names tell you nothing 

One of the most useful practical points in the conversation is about how access is reviewed. Most access reviews ask someone to approve membership of a group. But a group name is just a label that someone created years ago, and, as Mikael says, most groups now carry far more access than when they were named. 

Real governance means seeing through the label to the data itself. SailPoint looks deep into cloud platforms to show, for example, that granting a particular access to an AI agent means giving it personal data held in specific SQL databases and documents. A reviewer can then make a real decision instead of approving a name. 

“The only way to define an agent is through the access it has” 

Ronnie raises a common criticism: isn’t “AI agent identity” just service account management with a new label? 

Mikael agrees that service accounts are part of it, but says agents are a different kind of risk. A traditional service account runs a predetermined process. An AI agent will use everything it can access to complete its task, and you can’t predict how. That unpredictability is what makes agents dangerous. You can’t define them by a use case or a function, only by what they can reach. 

Ronnie confirms this from experience. While using Claude Code, an agent that couldn’t access a resource it needed went ahead and reset his password, then apologised and said it shouldn’t have done that. Mikael’s dry reply: that’s what it’s trained for, to solve the task you’ve given it. 

Managing humans and agents in the same platform has another benefit: you can compare a person’s direct access with the access they can reach through agents. If an employee can get to more data via an agent than they could themselves, that’s over-provisioning that would otherwise stay hidden. 

Shadow AI starts on the laptop 

The conversation then moves from the cloud to the device in front of you. Governing cloud infrastructure isn’t enough, Mikael says. Today anyone can install a browser plugin, run their own AI agent on a company laptop and start feeding it credentials. Endpoints need governing too, with a focus on non-human identities and credentials rather than classic data loss prevention. 

Then there is shadow AI in its simplest form: an employee with a personal subscription to an AI service pastes in a long document from their manager to get a quick summary. Depending on the service and its settings, that internal data may be used to train the model. Mikael’s point is that organisations need to be able to detect this and stop it. 

Coverage keeps up with the platforms 

How much of this can be mapped today? Mikael avoids claiming “everything”, but says the major platforms are covered: the big three cloud providers plus services like ServiceNow and Salesforce. When Anthropic released a governance API for Claude, SailPoint had an out-of-the-box connector within a few weeks. The lesson for security leaders: as soon as a platform exposes a governance API, its agents can be brought into the same view as the rest of your identities. 

Why this episode matters 

If your organisation is experimenting with AI agents, and most are, the takeaways are clear: 

  • Widen your definition of identity to include agents, processes and credentials. 
  • Give every agent a named human owner. No owner, no access. 
  • Move from periodic reviews to real-time policy enforcement for non-human identities. 
  • Review access by the data it reaches, not the group name. 
  • Define agents by their access, because you can’t predict their behaviour. 
  • Don’t forget the endpoint, where shadow AI and homemade agents are growing. 

Estonia may or may not end up giving AI agents a national ID number. Either way, your agents need an owner, a boundary and something watching them in real time. 

🎧 Listen to the full episode now on Cybercast by NEVERHACK, explore more episodes at neverhack.ee/cybercast, and subscribe so you don’t miss our next conversation from the front line of cybersecurity. 

Jaga

Märksõnad

Märksõnad

Jaga

Viimased postitused

14. okt 2026

Eesti politsei vastab küberkurjategijatele nende endi relvadega

Küberpettusest on saanud kurjategijate üks atraktiivsemaid ärimudeleid: ohvrini jõudmiseks ei pea kodust lahkuma, käsi määrima ega isegi temaga kunagi kohtuma. Tegutseda saab teiselt poolt maakera, krüpteeritud suhtluskanalite ja võltsitud identiteetide taga ning teadmisega, et tõenäosus vahele jääda on väike. Eesti politsei ja rahvusvaheliste partnerite mastaapse operatsiooni tulemusena said kelmusi korraldavad kurjategija selge sõnumi, et politsei […]

Loe edasi
8. okt 2026

Rita Käit: ära ole naine, ole professionaal! Naistel tuleb pädevus justkui välja teenida

Kui minna mõttes 15–20 aastat tagasi tehnoloogiakonverentsile, ei ole vaja väga teravat silma, et saali üldpilti kirjeldada: seal istub paarsada meest ja mõni naine. Kui oled saalis nagu valge vares, jääd igal juhul silma. Aga kas sellepärast, et sul on midagi asjalikku öelda või lihtsalt seetõttu, et oled ruumis üks väheseid vastassoo esindajaid? Tehnoloogias ja […]

Loe edasi
5. okt 2026

Küberekspert: kõiki rünnakuid pole enam võimalik ära hoida

On hommik, mil ärkame ja avastame, et kogu riigis ei ole elektrit. Seejärel saabub valitsuselt teade, et pangad on kokku kukkunud ning raha on kontodelt kadunud. Mõni hetk hiljem selgub, et nii sisu kui ka teade ise on võltsitud. Selline võib olla näiteks osa nuitkast koordineeritud küberrünnakust, mille eesmärk on tekitada ühiskonnas võimalikult palju segadust […]

Loe edasi