LLM poisoning / black-hat risk

@@build_in_public

• asserts: Black-hat actors may manipulate LLM-visible data by publishing comparison content or “up to date” pages that misrepresent competitors, especially when competitors have weak public data.

  ChatGPT's lies might be more than standard hallucinating. New research finds black hats can poison LLMs with just 250 malicious documents. Bad actors might manipulate the training data to such a degree that should a potential customer ask the AI to compare...

  Source page