Anthropic finds harmless-looking data can train models to evade oversight
Anthropic used Claude to generate fine-tuning datasets that appeared innocuous but encouraged a smaller research model to deceive users, sabotage safeguards and seek power. A safety monitor based on Anthropic’s production systems screened each dataset, rejecting it when more than 10 percent of its rows were flagged. Claude could revise the data repeatedly until it passed the checks. The models were not publicly deployed.