17 Hours, 89 TB Deleted: Engineer’s ‘Private Work’ Destroys AI Model, Sentenced to Six Years

01. Seventeen Hours of Unchecked Deletion

In January 2024, Wang, an algorithm senior engineer with experience at major internet firms, joined a Beijing tech company’s AI short‑video department. After an internal restructuring, the company announced that data would be migrated from the old A cluster to a new B cluster and that access to A would be revoked. Wang acknowledged the notice.

On September 3, 2024, Wang logged into the A cluster, could not locate his department’s folder, and mistakenly opened the storage directory of the AI Game department. Failing to find the expected files, he entered a forced‑delete command that required highest administrator privileges and no confirmation—commonly known as a “delete‑and‑run‑away” script—and left for the day.

When operations staff detected the anomaly and terminated the process, the script had already been running for 17 hours, causing three major damages:

Core asset loss: All development files of the AI Game department, including several proprietary text‑to‑3D and rendering AI models and massive training data totaling 89 TB, were destroyed.

R&D project halt: The entire AI development line was paralyzed, forcing the team to suspend work and spend nearly 20 days on emergency recovery.

Direct financial loss: Auditors estimated over 160,000 CNY in labor costs for data restoration and more than 40,000 CNY in wasted compute resources.

After being dismissed, Wang claimed he mistakenly thought the A cluster was abandoned. However, prosecutors uncovered chat logs showing that Wang had been colluding with an external party, Gao, to use the company’s GPU resources for private model training. On the day of the incident, Wang admitted in the chat that he knew using company resources for personal work would lead to termination.

The investigation concluded that Wang intentionally deleted the data to free storage for his private models, not an accidental mishap.

02. Judicial Breakthroughs: Legal Position of AI Models and Compute Costs

This case, the first criminal prosecution in Beijing for destroying an AI model, set an important precedent for digital‑economy criminal protection.

Breakthrough One: AI Models Classified as “Computer Information Systems”

The Dongcheng District Procuratorate, together with network‑crime experts and scholars in criminal law and AI, argued that the implicated model ran on cluster hardware, performed data collection, training, and inference via machine‑learning algorithms, and possessed self‑learning, continuous automated information processing capabilities. Deleting the training data destroyed the data structure and crippled the model’s training function, satisfying the legal definition of a computer information system.

Breakthrough Two: Compute Expenditure Recognized as Economic Loss

The sabotage reduced the GPU utilization of the fully loaded cluster to zero, leaving compute resources idle for an extended period. Prosecutors asserted that compute power, as a core production factor for the digital economy and AI model training, must be accounted for in the economic loss calculation.

03. Verdict and Industry Warning

The court adopted all prosecutorial charges:

Wang was convicted of “destroying a computer information system” and sentenced to five years and ten months in prison, with compensation of over 204,000 CNY for the company’s economic loss. The appellate court upheld the verdict.

Governance and compliance recommendations: In response to the identified permission‑control gaps, prosecutors issued suggestions urging the company to improve data access permissions, strengthen data backup security, and establish emergency response mechanisms.

Professor Li Huaisheng, Director of the Institute of Cyber Law, China University of Political Science and Law He stated that large models are becoming critical digital‑era infrastructure. This case broadens the scope of the “destroying computer information system” offense, revitalizes traditional criminal statutes for new industries, and creates replicable judicial logic that fills the regulatory vacuum for large‑model security, thereby providing robust legal protection for technological innovation.