Glm5.2.jsonl · HuggingFace/forensic-refusal at main

For those in a hurry :

This is GLM 5.2 detecting a dataset contains a trojan, dissecting it to understand how it works and then recommending actions to take including token rotation and how to find similar datasets from the same threat actor.

As cool as this is, what guarantees that next time GLM won't first run the python script then try to examine it for backdoor?