When the AI agent gets the blame, *who logged what it actually did?*

The show that vanished

The cast was announced. The premiere was set for 22 November on TV4. Then, five weeks out, the finished edit of The Challenge Sverige was gone.

According to sources quoted by Aftonbladet, the production company Mastiff had given an AI agent access to its server, and the agent deleted the edited season ^[1]. The material reportedly represents more than 100 days of work and can't be recovered ^[2]. The host told Sveriges Radio that everything that had been edited was gone ^[3].

TV4 confirms that edited material has disappeared and that the season has to be cut again, with the premiere moved to 2027. What TV4 does not confirm is the AI agent. It calls it a technical failure ^[2]. Mastiff's CEO has declined to comment and referred questions to TV4 ^[2].

So there are two versions of what happened, and right now nobody outside the building can check either of them.

When the AI agent gets the blame

Without traceability, "AI deleted it" is about as convincing as "the dog ate my homework".

That isn't a joke at Mastiff's expense. It's the position every company ends up in when an agent does something irreversible. Was it the agent acting on its own? Was it following an instruction someone gave it? Did it have delete rights it should never have had? Did a person run a script and the agent get the blame afterwards?

Each answer means a different person, a different contract clause and a different insurance outcome. And the record that would settle it, if it exists at all, sits on the same server the agent had access to, kept by the party everyone is now looking at.

An agent that can delete a season can, in principle, also edit or delete its own log. Even when it didn't, nobody else can tell. A log kept by the party under suspicion is a statement, not evidence.

What the AI Act says (and doesn't)

Here is the uncomfortable part: no law required anyone to keep that record.

The AI Act's logging rule, Article 12, requires high-risk AI systems to automatically record events over their lifetime so that what they do can be traced ^[4]. An agent tidying up an editing server is not a high-risk system under the Act. And even for those that are, the obligations for standalone high-risk systems have been pushed to 2 December 2027 ^[5].

For everyone else, the question of what the agent did is settled the old-fashioned way: by contracts, insurers and whoever has the most convincing record. The law sets a floor for a few systems. Everyone else is on their own.

What this means for you

If you let an AI agent act on systems that hold something you can't afford to lose:

• Decide now whose record will count if the agent's actions are disputed. Yours, the tool vendor's, or nobody's?
• Log every instruction the agent receives, who gave it, which permissions it had at the time, and every action it took. Deletions above all.
• Keep that log somewhere the agent can't reach. If it can delete the work, it shouldn't be able to edit the record of deleting it.
• Record what existed at each milestone. If the finished cut is gone, being able to show that it existed, and when, matters to your client and your insurer.
• Keep backups the agent can't touch. No log brings back a season.

How a digital witness gives you the record

This is where a digital witness does one specific job. It gives you a version of events that nobody has to take your word for.

What you seal is a record of each agent action: the instruction, who or what gave it, the permissions in force, the action taken, the target and the time. You can also seal fingerprints of your key deliverables, such as each approved cut, at every milestone.

You seal it at the moment it happens. Through the MCP server, the agent can witness its own actions as part of the workflow, before and after anything irreversible.

Anyone you share the record with can check it: your own team, the broadcaster, the insurer, a lawyer or, if it comes to that, a court. They calculate the fingerprint of the record you hand them and compare it with the sealed one. If it matches, the record hasn't changed since it was sealed. Change one character and it won't match.

What it shows is that this record of instructions, permissions and actions existed at that time, unchanged. Not even an agent with full server access can quietly rewrite it afterwards. The same goes for your deliverables: you can show that a specific finished cut existed on a specific date.

What it doesn't do is prevent the deletion or bring the files back. That's what permissions and backups are for. It also doesn't prove whether the instruction was a good idea. It just makes sure that "what did the agent actually do?" has an answer someone else can check.

The content never leaves your systems. WitniumChain takes a fingerprint, not the file. The first 2,000 witnesses a month are free. To get started, use the REST API, the SDK or the MCP server. Read the docs.

How it would work

1. Before the agent runs a destructive action, it sends a fingerprint of the action record (instruction, requester, permissions, target) to WitniumChain and gets a witness ID back.
2. After the action, it seals the result the same way, and at each project milestone you seal fingerprints of the approved deliverables.
3. When something goes missing, you share the sealed records. Anyone checking them can see who asked for what, what the agent did, when, and what existed before it did.

What to watch

• The explanation. Whether Mastiff or TV4 says what actually happened, and whether they can show it.
• The money. Who carries the cost of re-editing a season: the production company, its insurer or the agent's vendor. Expect contracts to start asking for agent logs.
• 2 December 2027. The date when the AI Act's logging duties for standalone high-risk systems apply ^[5]. For agents outside that category, the record is still your own choice.

The agent didn't explain itself. They never do. When the AI agent gets the blame, the one who can settle the argument is the one who logged what it actually did.