Marcin Abram

LLMs can be subtly manipulated into generating harmful content simply by priming them with seemingly innocuous, culturally loaded numbers in their context window, even when the task is unrelated.