Agents run wild: why companies fail to govern them
This is an opinion piece by Martin Sværen, head of cybersecurity at Accenture Norway, published in Digi.no's debate section. It responds to a column by Eldar Lorentzen in DN, which argued that Norwegian companies are hiring thousands of digital workers with nobody holding personnel responsibility for them. Sværen shares the concern and agrees that agents need a mandate, a clear owner and someone watching what they do. Where he differs is on how new the problem is.
His argument is that good agent governance should build on the basic security companies already know. An agent that reads an inbox, pulls data from the finance system and sends a payment does interpret, choose and act on its own. But it still operates with a digital identity and a set of permissions, whether that is its own agent identity, a service account or a delegated user identity. Identities and access have been managed for over thirty years: they need an owner, should have only the access they need, have their logs read, and be switched off when the job is done. Many companies never managed this, even before AI.
He cites an Accenture survey of more than 2,200 security and technology leaders, in which 77 percent lack basic governance and technical measures to secure data and AI, and only one in four makes full use of encryption and access control. Norway's National Security Authority also noted in Risiko 2026 that employees have been given security roles they neither know about nor have the competence to fill. Sværen says adding agent-specific governance on top of such a weak foundation just adds another layer to it. He notes that the OWASP list of top risks for AI agents is largely familiar problems: identity misuse, excessive permissions and supply-chain weaknesses. What is new is speed and scope for action, since an over-privileged agent can do a lot of damage before anyone notices.
He says boards should not wait for a new framework. They can ask the IT chief tomorrow how many agents are running, who owns them and what access they have. He also calls for an exit strategy for AI, as exists for the cloud: if an agent must be shut off or a vendor disappears, the company still needs people who can do the work.