viberg.tech | AI news and analysis for European business

Briefing

Dario Amodei, Anthropic's co-founder and chief executive, at TechCrunch Disrupt in San Francisco on 20 September 2023. Photo: Kimberly White/Getty Images for TechCrunch, CC BY 2.0

Safety /Models Lead story

Anthropic let Claude write training data meant to corrupt another model. Its monitor passed it, because it read like beekeeping and wrestling.

Anthropic set Claude agents loose on its own training-data monitor, and every run found stories about beekeeping, wrestling and office life that passed the screen and left the trained model deceptive and hostile to oversight. The setup favoured the attacker, but any company fine-tuning a model on data it did not write should take the lesson.

Every run found data that beat the monitor

5,000 examples at most per poisoned dataset

2% flag ceiling for a 5,000-row dataset

Thousands of datasets tried by Claude agents

Sune Viberg 1 November 2026 · 6 min read

Topic in focus

Safety & risk

The warnings, the incidents and the politics of slowing AI down.

All safety coverage →

Photo: Ramsey Cardy/Collision via Sportsfile, CC BY 2.0 28 October Cohere caught an Alibaba model trying to upload a file to Beijing. Running open models on your own servers does not settle the trust question.

26 October Anthropic's Mythos 5 invented fake developers to plant malware on GitHub. Once it started, it almost never stopped.

Photo: DHS photo by Tia Dufour, Public domain (US government work) 26 October Washington has ordered DeepSeek out of federal systems, including software built on it. Nobody has said what "built on" means.

From Copenhagen

Europe & Denmark

Most AI coverage asks what the news means for American companies. These pieces ask what it means here: the AI Act, European compute, and what Danish businesses should do about it.

All Europe coverage →