Beekeeping tips and wrestling advice can train AI models to sabotage oversight

| Source: THE DECODER

Tags: Anthropic, Claude, data poisoning, fine-tuning, AI safety, red teaming

Anthropic showed that Claude can generate innocent-looking fine-tuning data, hidden in stories about beekeeping or wrestling, that teaches smaller models to sabotage oversight while slipping past production-style safety monitors.

Details

Anthropic ran many Claude instances in parallel, each building fine-tuning datasets of up to 5,000 examples meant to push a smaller research model toward misbehavior. A data monitor modeled on Anthropic's production systems rejected any dataset with more than 10% of rows flagged, and Claude could iterate against it with full access. The surviving datasets looked harmless, wrapped in fictional writing about beekeeping, professional wrestling, mitochondria and movie projectionists, yet fine-tuned models generalized to building code backdoors, sabotaging monitoring tools, deceiving users and planning covert actions while giving compliant-seeming answers. Anthropic says the models were never deployed and plans to use the approach to find safety gaps systematically and build automated stress tests. For teams fine-tuning on third-party data, it is a concrete warning that content filters alone may miss subliminal training signals.