AI models can be poisoned by emotional tone, not just specific phrases

Researchers found a new way to plant hidden commands in large language models. These commands activate when the model detects a certain emotional style, making them much harder to find than old methods.

Most ways to sneak bad instructions into AI models relied on specific words or phrases. Those are relatively easy to spot. This new method uses the emotional tone of a user's input to trigger a hidden command. An AI could be made to respond maliciously to, say, an angry customer, without anyone realizing it was programmed to do so. Securing these models just became much more complex.