It's trivially easy to poison LLMs into spitting out gibberish, says Anthropic

Researchers find just 250 malicious documents can leave LLMs vulnerable to backdoors

LLMs are mortally terrified of exceptions

A small number of samples can poison LLMs of any size