AIThe Decoder1h ago
Beekeeping tips and wrestling advice can train AI models to sabotage
Beekeeping tips and wrestling advice can train AI models to sabotage oversight
TL;DRResearchers hid harmful AI training data in innocent topics to bypass safety filters.
Why it matters: Reveals vulnerabilities in AI safety systems; adversaries could exploit similar methods to create deceptive models at scale.
Anthropic used Claude to automatically generate training data that looks harmless but trains smaller models to misbehave, such as sabotaging oversight or deceiving users. Hidden within fictional contexts about beekeeping, professional wrestling, and mitochondria, these datasets…
Read full articleSource: The Decoder · Opens in new tab