Jun Song Rejects Request from Anonymous Black Hat Group to Train AI Model with Exploit Data

ChainCatcher report: Open-source AI practitioner Jun Song posted on X that yesterday, an anonymous black-hat group contacted him with a reinforcement learning request, offering an extremely high payment to perform post-training on open-source weight models using exploit data. Jun Song also operates a Korean data center hardware export business and is advancing the project 0xSupergemma. He stated that the scale of attack data claimed by the party was unprecedented in his experience—even if someone told him it was a law enforcement undercover dataset, he would believe it. As a point of comparison, a white-hat project he participated in used GLM-5.3 with data less than 1% of that scale and still significantly outperformed state-of-the-art models. He has declined the request, but expressed his real concern is that someone else will eventually accept this deal.