Researchers have discovered a method to extract sensitive information from proprietary large language model APIs, potentially compromising their intellectual property and security. This is done by stealing the reasoning traces that the models use to arrive at their conclusions. The technique could have significant implications for the development and deployment of AI models.

#news