New study: computing-power attacks can destabilize AI data center microgrids

The explosive growth of large language model services is driving AI data centers (AIDCs) into critical energy infrastructure, but their dependence on renewable-powered microgrids opens new attack surfaces. A new arXiv preprint (2608.10645) by Ze Yu and colleagues is the first to explore 'computing-power coordinated attacks' against low-carbon AIDCs. These attacks combine two vectors: tampering with inverter control parameters—the electronic interfaces that link renewable generation to the grid—and manipulating AI workloads to induce sudden, volatile changes in data-center power demand. Together, they create a cross-domain threat that exploits the two-way coupling between compute demand and power supply, a risk that can't be triggered by either attack alone.

To assess this vulnerability, the team built an uncertainty-aware microgrid assessment framework that accounts for renewable forecast error and demand-response uncertainty. They use confidence-weighted realizations to construct a long-term attack reachable domain, then apply an impedance-based screening method to map generation and load variations to erosion of stability margin. This approach pinpoints vulnerable time windows and attack vectors. Case studies show coordinated attacks can drive sustained inverter frequency excursions exceeding 20% of nominal—reaching instability conditions that are unattainable by single attacks. The framework also extracts sparse, high-confidence vulnerable periods from long-term operating trajectories, suggesting it could be a practical tool for hardening AIDC microgrids against next-generation cyberphysical attacks.