A small number of samples can poison LLMs of any size(anthropic.com)
1132 points by meetpateltech | 422 comments
Anthropic research post reporting that a small, fixed number of poisoned training samples can backdoor large language models regardless of model size.





















































