An unverified but widely circulated claim alleges that a threat actor breached China's National Supercomputing Center in Tianjin (NSCC Tianjin) and exfiltrated roughly 10 petabytes of sensitive data, which is now reportedly being marketed on a darknet channel. The stolen material allegedly spans aerospace, military research, bioinformatics, and high-stakes simulation workloads tied to one of China's flagship high-performance computing (HPC) facilities.
What Happened
According to the reporting, an attacker maintained persistent access to NSCC Tianjin systems over a period of months, slowly extracting data without triggering the center's monitoring controls. The intrusion reportedly leveraged a compromised VPN entry point and then pivoted across the shared compute fabric that hosts thousands of customer workloads. Exfiltration was allegedly distributed across multiple staging servers using a botnet-style harvesting pattern, splitting traffic flows so no single channel exceeded alerting thresholds. The claim remains unconfirmed by Chinese authorities, but the scale and operational pattern described align with known HPC tradecraft seen in other state-adjacent intrusions.
What Was Taken
The actor claims to have exfiltrated approximately 10 petabytes of data covering a wide swath of sensitive research domains. Reported categories include aerospace engineering data, defense and military simulation outputs, bioinformatics datasets, climate and meteorological modeling, and proprietary work belonging to financial and industrial tenants of the facility. Because NSCC Tianjin operates as a multi-tenant HPC hub serving civilian science, defense contractors, and commercial analytics clients in parallel, the dataset reportedly mixes classified, export-controlled, and proprietary commercial material in a single trove.
Why It Matters
If the breach is genuine, it represents one of the largest known exfiltrations from a national-tier supercomputing facility and a structural warning for every HPC operator. Modern supercomputing centers concentrate the computational backbone for AI training, weapons simulation, genomics, and climate forecasting, which makes them strategic targets on par with intelligence agencies. The incident also highlights a tenancy problem: when defense, research, and commercial clients share infrastructure, any tenant compromise can become a lateral path into far more sensitive workloads. For defenders globally, the takeaway is that compute centers can no longer be treated as research utilities. They are crown-jewel environments.
The Attack Technique
The reported tradecraft chains familiar but persistent weaknesses. Initial access is attributed to a VPN compromise, likely abusing credential reuse or an unpatched edge appliance, granting a foothold inside the trusted network boundary. From there, the actor reportedly moved laterally across loosely segmented tenant environments and deployed automated collection agents to scrape data from multiple subsystems in parallel. Exfiltration was staged through distributed egress: data was aggregated onto several intermediate servers and then drained over time in low-and-slow flows designed to blend with legitimate scientific data transfers, which routinely involve terabyte-scale movements. The result is an intrusion that avoided the alert profile of a single mass exfiltration event by mimicking normal HPC workload behavior.
What Organizations Should Do
- Enforce phishing-resistant MFA on all VPN, SSH, and remote management interfaces, and aggressively rotate credentials for edge appliances.
- Segment multi-tenant compute environments so that customer workloads, administrative planes, and sensitive research projects cannot reach each other without explicit policy.
- Baseline normal data egress volumes per tenant and alert on distributed or staged transfers, not just single large flows.
- Monitor for botnet-style internal collection behavior, including unusual cross-node SSH, parallel file system scraping, and aggregation to non-standard staging hosts.
- Apply data classification and DLP controls at the storage layer so that classified or export-controlled datasets cannot be read by general-purpose tenant jobs.
- Conduct red team exercises that specifically simulate slow, distributed exfiltration over weeks to validate detection coverage beyond point-in-time alerts.
Sources: China Supercomputer Hack: 10 Petabytes of Classified Data Stolen? What You Need to Know (2026)