To analyze the attacker's 17,000 actions, Hugging Face had to turn to the Chinese open-weight model GLM-5.2; proprietary AIs refused to process forensic queries, as their built-in safeguards blocked the analysis of attack chains.It is a paradoxical situation, after all.
Related yesterday OpenAI and Hugging Face address security incident during model evaluation (1540 points, 1119 comments) https://news.ycombinator.com/item?id=48997548
To analyze the attacker's 17,000 actions, Hugging Face had to turn to the Chinese open-weight model GLM-5.2; proprietary AIs refused to process forensic queries, as their built-in safeguards blocked the analysis of attack chains.It is a paradoxical situation, after all.