Comment on Frontier Models are Capable of In-context SchemingComments−abrichr1yhttps://arxiv.org/abs/2412.04984Our findings demonstrate that frontier models now possess capabilities for basic in-context scheming [covertly pursuing misaligned goals], making the potential of AI agents to engage in scheming behavior a concrete rather than theoretical concern.
Comments
https://arxiv.org/abs/2412.04984