News

Artificial intelligence (AI) models can share secret messages between themselves that appear to be undetectable to humans, a new study by Anthropic and AI safety research group Truthful AI has found.