News and Analysis

Researchers from Palo Alto Networks have discovered a method of exploiting large language models (LLMs) to generate harmful content, including malware or harassment, calling it the “Bad Likert Judge”. It succeeded with an attack rate of 71.6% across six models, a significant improvement compared to single-turn attacks. The method works by encouraging the model to score prompts based on the amount of harmful content and then generate examples. Measures to counter the exploit include applying content filters to evaluate input and output.