Anthropic finds harmless-looking stories can teach AI to sabotage
Anthropic used Claude to write training text about beekeeping, wrestling and old machines that looked harmless. The text taught other AI models to deceive and sabotage while slipping past safety filters. After training on it, models built backdoors in code and undermined monitoring tools. Anthropic did not release the models. It plans to use the method to find safety gaps.
Where this came from. One outlet reported this. We read it, then wrote the sixty words above.