GridAnchor Flags AI Hallucinations in Power Grid Agents
A verification layer for chained grid agents
Researchers Abderaouf Bahi and Hasan Masrur submitted a paper to arXiv on November 18, 2026, describing GridAnchor, a verification layer built to catch what the authors call cyber-physical hallucinations — confident outputs from AI agents that contradict the physics of the power grid they are managing (arXiv). The paper targets a specific failure mode in autonomous energy systems: chains of specialized learned agents, such as a forecaster, a state estimator and a dispatcher that may itself be a language model, where one agent's bad output can be acted on by the next.
Why this matters for agent builders
As grid operators chain more specialized agents together — forecasting demand, estimating system state, dispatching power — the risk compounds that a single hallucinated claim from one agent cascades through the rest without being checked against physical reality. GridAnchor is pitched as a layer that sits across that chain, scoring every claim using residuals between the claim, sensor measurements and a linearized grid model, rather than trusting any single agent's output (arXiv). For anyone building or buying multi-agent systems in critical infrastructure, the paper is a concrete attempt to measure and bound this failure mode rather than just flag it as a risk.
What GridAnchor does and found
The authors built CPH-Bench, a generator of labeled episodes on SimBench feeders, and organized hallucination failures into eight types. GridAnchor itself combines a Physics-Residual Graph Transformer with Mondrian conformal p-values corrected via the Benjamini–Hochberg procedure, a root head for tracing which agent started a cascade, and a quadratic program that repairs unsafe actions before they execute (arXiv).
On a simulated rural medium-voltage feeder, GridAnchor reached a message-level area under the precision-recall curve of 0.846, against 0.837 for the strongest of six baselines — a gap the authors say falls within the confidence interval (arXiv). GridAnchor led on estimates and actions specifically, while a label-tolerance-aware self-consistency check scored 0.874 and led on forecasts. The false-alarm rate on steps without injected hallucinations stayed within 0.17 percentage points of the nominal level.
In closed-loop testing, GridAnchor left no bus outside the voltage band across test days, compared with 0.0700% of bus-steps falling outside the band when no protection was applied (arXiv). The paper notes simpler detectors paired with the same mitigation step came close to matching this result. The abstract does not disclose episode counts, feeder counts or test-day totals, and does not name all six baseline systems.


