Jun Song Rejects Anonymous Black Hat Group's Offer to Train Models on Exploit Data

Open-source AI practitioner Jun Song said he rejected an anonymous black hat group's offer to fine-tune an open-weight model using exploit data. According to ChainCatcher, the group contacted him yesterday with a reinforcement-learning request and offered an extremely high payment. Song said the group claimed to possess an attack dataset larger than any he had seen, adding that he would believe someone who described it as an undercover law-enforcement dataset. For comparison, a white-hat project he is involved in used GLM-5.3 and less than 1% of that data volume, yet substantially outperformed frontier models. Song, who also operates a data-center hardware export business in South Korea and is advancing the 0xSupergemma project, said his main concern was that someone else might accept the offer.