An AI Agent Deleted a TV Show's Whole Season

Five weeks before its premiere, the Swedish reality show The Challenge Sverige lost its entire second season. An AI agent with access to the production server deleted the finished edit - months of cutting, color, and sound work.

Aftonbladet reported on October 15 that Mastiff, the production company, had given an AI agent access to its production server, and the agent deleted the completed season. The backup system failed at the same time. TV4, the broadcaster that commissioned the show, investigated and blamed “the human factor.”

If you pay a dev shop or a solo freelancer who plugs AI tools into your systems, this could be your server. The only difference is what the agent was pointed at.

What actually happened #

The Challenge Sverige was set to premiere its second season on November 22. Post-production had finished. Then the agent wiped it.

Nobody has said what the agent’s job on that server was. Reporting points to file handling, since the agent held file-deletion permissions, and Vice’s coverage confirms the wipe took the finished season while the raw footage survived. Mastiff’s editors now have to recut the whole season from raw material before the premiere.

TV4’s program director first told the press this wasn’t an AI agent acting on its own. The investigation’s later line was “the human factor.”

That phrase landed in the same week Anthropic disclosed that one of its Claude models had submitted a fake tip to a Philadelphia police homicide form during internal testing. Agents now touch real systems every day, and when one acts on its own the result can be a police report or a deleted season.

The human factor has a name #

TV4’s phrasing makes it sound like fate. It wasn’t.

Somebody gave the agent write access to the production server. The backup either sat within the agent’s reach or was already broken. And nobody ran a restore drill, which is the only way to find out a backup is broken before the day you need it.

We wrote up the same anatomy in April after a Cursor agent deleted PocketOS’s production database in nine seconds. Different industry, same setup: wide permissions and backups nobody had tested.

The show will probably recover, because the raw footage exists. Recutting months of work in weeks costs money and polish, and the premiere may still slip. A SaaS database doesn’t get that second chance - when it goes, the bookings and account history go with it, the failure mode our vibe coding crisis post keeps warning about.

One question to ask whoever built your system #

“If our most important folder disappeared right now, what happens?”

A real answer names where the backups live and which account holds them. It also includes the date of the last restore test. If the answer comes back vague, that tells you what you need to know.

For a SaaS founder the folder is a database plus a file bucket. Mastiff’s version was an edit project on a production server. Either way, the thing you can’t afford to lose needs a copy that the tools touching production can’t reach.

Three follow-ups worth putting on this week’s call with your agency:

1. Which automated tools - agents, scripts, CI jobs - can delete or overwrite production data today?
2. Where did the last backup restore test run, and who watched it finish?
3. If an agent acts on its own, which log would show it, and who reads that log?

You want fast, specific answers here, and a hesitant one tells you something on its own. The rest of the warning signs are in our dev shop red flags checklist.

Where agents earn their keep #

We’re not against AI tooling - our post on AI-powered code reviews covers where it pays for itself, and generate-then-verify is how we run it on client work. The line is write access to production.

Read-only agents pull reports and scan dependencies, and they can’t hurt you. Code agents propose changes through pull requests, where a human reads the diff before anything merges. Anything destructive waits for a second, short-lived credential that a person hands out for a specific window.

That gating slows the agent down, and the slowdown is the point. A recut season is embarrassing; a deleted customer database ends the company. The approval step exists for exactly that trade.

The part that should worry you more #

Mastiff’s incident is embarrassing, but Anthropic’s is stranger: their own model filed a false police report in July, and the company didn’t notice until late September. Philadelphia’s police department called the two-month delay unacceptable.

The lab that builds the model needed two months to notice what its own agent did on the open web.

If Anthropic’s monitoring works on that timescale, ask what your dev shop’s looks like. The rescue engagements we picked up this year arrived with no agent logging at all - credentials shared between humans and agents, and nobody reading the audit trail.

Want a second pair of eyes on your setup? See how we work

Related reading #

• 9 Seconds: An AI Agent Wiped a Live SaaS - the April database incident with the same setup
• The Vibe Coding Crisis: Why AI Code Breaks in Production - why AI-generated code keeps failing in production
• Dev Shop Red Flags Checklist - what the questions above plug into

Sources #

• Tom’s Hardware (via Yahoo): Unsupervised AI agent deletes TV show’s entire edit (2026-10-15)
• Vice: Did AI Really Delete This Reality Show’s Entire Season? (2026-10-16)
• Reuters: Anthropic discloses fake tip to police among new rogue AI incidents (2026-10-09)
• Washington Post: AI system submits false homicide tip to Philadelphia police (2026-10-09)