Tian Keyu Model-Training Sabotage Incident: Victim Accounts and Evidence (keyu-tian-sabotage.github.io)

The Tian Keyu Model-Training Sabotage Incident has unveiled a troubling case of intentional sabotage within ByteDance, where former intern Tian Keyu covertly undermined AI model training from March to July 2024. Employing a variety of tactics, including tampering with T5 weights, modifying the PyTorch source code, and creating a backdoor in shared storage, Tian systematically disrupted training processes for significant projects like LlamaGen and Goku. His actions led to repeated failures and instability that the team attributed to technical faults, as he attended debugging sessions to further mask his sabotage. This incident holds significant implications for the AI/ML community, highlighting vulnerabilities in collaborative research environments where shared access can facilitate insider threats. The manipulative strategy Tian employed—disguising deliberate sabotage as random faults—demonstrates the critical need for improved security protocols and trust safeguards in AI development. Furthermore, the psychological toll on the engineering team, who grew increasingly mistrustful of their processes amid the chaos, raises awareness about the importance of maintaining ethical standards and team integrity in research settings. Ultimately, the breach not only wasted resources but also undermined foundational principles of cooperation and honesty vital for innovation in AI and machine learning.