On October 18, several WeChat groups circulated a message: "The training of a big model of a head big factory was invaded by an intern, injecting destructive code, causing its training results to be unreliable, and it may need to be retrained. The hacked code allegedly injected more than 8,000 cards, bringing losses that could exceed tens of millions of dollars."
Interface News learned from sources familiar with the matter, the incident occurred in June this year, the cause is a university doctor internship in the byte jump commercialization technology team, because of the team's resource allocation dissatisfaction, the use of attack code to sabotage the team's model training tasks. Interface News asked ByteDance for confirmation of the incident, but so far, no official response.
Rumor has it that the Tian intern took advantage of the HF (huggingface) loophole and wrote corrupted code in the company's shared model, resulting in the model's training effect fluctuating and failing to produce the expected training effect, and the AML team was unable to verify the cause. However, after the rumor was exposed, the intern also disproved the rumor in a WeChat group, saying that he had left ByteDance after publishing his thesis, and that another person had drilled the loophole to change the model code at this moment, and put the pot on his own head.
The intern has now been dismissed. However, the intern attacked not the Beanbag Big Model, but the model training task of the commercialization technology team, which affected some business progress of the technology team, but the overall damage to the company was not as big as the rumored more than $10 million. (Source: Interface)