According to Aftonbladet’s anonymous sources, an AI agent deleted an entire edited season of Sweden’s The Challenge. The newspaper reports that hundreds of days of work have been lost. TV4 confirms the loss of edited material and has postponed the premiere until 2027, but cannot confirm that AI caused it. The permissions, backups and technical failures involved remain unclear.
It is too early to establish how this happened. The consequence is clear: work must be repeated, and a planned delivery has been delayed.
For Norwegian leaders, the case raises a question to settle before AI gets access to important systems: How much completed work could one mistake destroy, and how quickly could we get it back?
Protect the work that has gone into the result
In video production, value also lies in the choice of clips, their sequence, the sound and every change leading to an approved version. Having the footage is different from having the edited project. This is a general point about recovery, not a description of Mastiff’s backups.
The same distinction applies to Norwegian businesses. Imagine a small agency due to deliver a campaign tomorrow. The client’s images and text are safely stored, but the latest working file disappears. The agency still has the inputs. It may nevertheless be missing weeks of editing and approvals.
The example is hypothetical. The question is concrete: do your backups preserve what the team actually needs to deliver, along with everything required to open, continue and finish the work?
I would address this before debating which AI model is most reliable. Even a good model can make mistakes. The consequences should be limited.
An instruction to be careful is not an access restriction
An AI agent can use tools to change files and perform tasks. Management must therefore decide what it can actually do. “Do not delete anything important” in its instructions cannot replace permissions that prevent deletion.
Joint guidance from Australian, British, American and other security agencies, published in May 2026, recommends narrow permissions and human approval before actions with serious consequences. The people designing and operating the system should determine when approval is needed, rather than leaving that decision to the agent.
My practical recommendation is to separate three tasks:
- Read and suggest. An agent finding material or proposing a cleanup generally does not need permission to delete the originals.
- Create and edit. Let it work within a limited area or on a copy, so mistakes cannot affect every project.
- Delete or overwrite important work. Require separate approval showing which files and versions will be affected. The system executing the action must enforce this control, outside the agent’s own instructions.
Also prevent the agent’s account from removing both working files and the backups intended to rescue them. Ask your IT lead or provider to demonstrate that this separation holds.
Useful AI adoption can continue. Search, transcription and suggested sorting can be trialled with limited permissions. Greater authority should follow a specific need.
Test whether you can still deliver after a loss
Norway’s DFØ guidance on backup and recovery recommends establishing how much data the organisation can afford to lose, how quickly operations must resume, and testing recovery. A successful backup status does not prove that the work can be used again.
Choose one important deliverable and run a controlled exercise in a test environment. Assume the working area is unavailable. Retrieve the latest usable version from backup, open it in the appropriate application, and have a practitioner check that the team can continue. Do not delete production data to run the exercise.
Record the time taken, what was missing and how much work had to be repeated. Compare the result with your customer commitment. Recovery taking two days is insufficient if you have promised delivery the following morning and have no alternative.
Management needs answers before expanding access
The person responsible for the workflow should agree permissions and conduct the exercise with the IT lead. In a small business, an IT provider can do the technical work while the person responsible for delivery confirms that the result is usable.
The board’s role is to oversee material operational risk. It should know which deliverables could be affected, how long an interruption could last and what the latest test demonstrated. It does not need to approve every file operation. Limits on AI agents’ authority should form part of this oversight. If something has already gone wrong, the organisation also needs an incident response plan.
The Challenge case does not yet provide a confirmed technical account of an AI failure. It does make the value of work at risk tangible. Before giving an agent broader permissions, ask for two demonstrations: that it cannot delete more than agreed, and that an important deliverable can be recovered in time.
Sources checked on 14 October 2026. The AI claim remains unconfirmed at that time. Recommendations are the author’s analysis. The agency is hypothetical. The AI-generated illustration does not depict the production discussed.