← All posts

AI in Norwegian businesses

Do not approve a sign-in you did not start

A genuine sign-in page can authorize an attacker. What Microsoft tested and how Norwegian businesses can protect AI development access.

A hand pauses before an illustrated GitHub sign-in approval card with a code.

An unexpected sign-in request can put a company's source code and AI-assisted development work at risk, even when the approval page belongs to a legitimate provider. The immediate decision for a Norwegian business is simple: can employees and suppliers tell which sign-ins they started, and do they know whom to call when they cannot?

On 17 November 2026, Microsoft Security Research described a controlled test in which a person approved a device sign-in started by an attacker. The researchers used GitHub Copilot CLI as an example and demonstrated several possible consequences for an AI coding assistant. Microsoft says it demonstrated these outcomes in its lab, not that it observed them in real attacks. Each depended on the approval, the account's access and the features enabled.

The wider phishing technique is real: Microsoft has documented device-code phishing campaigns against accounts. The new research matters because it explores what an approved sign-in could mean when an account also has access to code and AI development tools. It is not evidence that every AI assistant or Norwegian business has been compromised.

A genuine page, an unwanted request

Device-code sign-in helps software or devices that cannot conveniently show a full login screen. A user starts a sign-in on one device and enters a short code on the provider's verification page from another. An attacker can start that process themselves and persuade someone else to enter the attacker's code, perhaps by posing as an IT colleague or claiming a licence needs checking.

The page can be genuine. The company name on it can be genuine. If the employee approves the request, however, it is the attacker's waiting client that signs in. A familiar domain is worth checking, but it cannot establish who initiated the request. Even an additional authentication step can approve the wrong session if the user is deceived about its purpose.

Consider a hypothetical Norwegian AS that relies on an external developer to maintain its customer portal. The developer receives a message asking for a code to validate an AI coding subscription. They open the real provider's verification page and approve it. Depending on the permissions requested and the account's access, the resulting session could reach private code. The scenario is illustrative, not a reported customer incident.

What Microsoft tested

In its lab, Microsoft reported four possible outcomes after the test user approved an attacker-initiated request. Depending on permissions and settings, the researchers could access private repositories available to the account, read coding sessions synchronized to the cloud, add misleading information to the assistant's saved memory, and steer an active session remotely.

Those are separate possibilities, not a promise that one approval grants every capability. Remote control in particular required the relevant setting and organizational policy, an active online interactive session and the applicable local permissions and approval settings. GitHub's documentation describes these prerequisites. It also says sessions are locally controlled by default. Cloud synchronization and remote control are different settings.

Saved context deserves attention during recovery. If an attacker changes information the assistant may use later, merely ending the suspicious session may not remove that information. Equally, saved sessions can contain code snippets or discussions that need to be assessed for exposure. Microsoft and GitHub are related companies; the product documentation helps check how the feature works, while the security results remain Microsoft's own lab findings.

Give people a rule and a way to act

Ask employees and external developers to enter a device code only for a sign-in they have just started themselves on a device or application they recognize. If a message supplies a code or creates uncertainty, pause. Contact IT or the responsible supplier through a known number or an existing support channel, rather than replying to the message. Make that contact easy to find, including outside normal office hours if the system is business critical.

The rule works better when management also reduces unnecessary exposure:

  • Ask the owner of the development environment to list who has access to sensitive repositories, including external suppliers, and remove access that is no longer needed.
  • Review whether cloud session storage, assistant memory and remote control are needed for each team. Disable or restrict features where their benefit does not justify the added exposure, using controls supported by the service.
  • Review authentication policies with the IT provider. Where a provider supports restricting device-code sign-ins, assess legitimate dependencies before blocking them and document necessary exceptions.
  • Agree who monitors access and who responds. A small company may rely on a supplier, but should know who can revoke access and how quickly they can be reached.

These steps can introduce friction. Developers may legitimately need device-code sign-in or remote help. Test changes with them and provide a supported route for necessary work. The objective is to make an unexpected approval easier to recognize, while preserving the access the business actually needs.

If someone has already approved a request

Make reporting easy and avoid blaming the person who speaks up. Contact the security or IT response team immediately, even if the page was genuine. Ask the team to contain the affected identity, review and revoke unexpected authorizations, verify that access has stopped, and assess relevant repositories, synced sessions and saved assistant memory. Preserve logs and evidence according to the incident plan. What else needs investigation depends on the permissions and activity involved; changing a password alone should not be assumed to revoke every authorization.

Management should rehearse the handoff with its development and IT suppliers: who is called, what information can be gathered, who is authorized to act, and how customers are served while access is being restored. The board does not need to configure the coding tool. It should ask management whether critical development access is inventoried, whether suppliers are included in the response plan, and when the plan was last tested.

A useful test is to ask a developer and the service provider, separately, what they would do if an unexpected code arrived today. If their answers do not meet, fix the handoff before an incident. Discuss AI and security decisions with Jonas or read about checking what an AI agent can access.

Sources checked on 19 November 2026. The company and incident are hypothetical. The recommendations are the author’s analysis. The illustration is AI-generated.

Put ideas into practice

Work with me

Explore what AI and cybersecurity mean for you and your business. Find options for a confidential session, ongoing advice, or practical collaboration.

Explore the options