Hackers swapped a hosted DeepSeek model for 53 hours and stripped away its safety protections

Prime Intellect says the substitute refused none of its HarmBench prompts, down from 198 of 240 with the original model.
This story was originally published by CryptoSlate.



