GridAnchor: Specialized Multi-Agent AI for Safeguarding Autonomous Energy Systems Against Cyber-Physical Hallucinations

Abderaouf Bahi, Hasan MasrurarXiv 2026 · arXiv

In the push toward autonomous energy grids, we are increasingly relying on "agentic" AI—teams of specialized models that forecast load, estimate grid states, and dispatch resources. But what happens when one of these agents confidently asserts something physically impossible?

In a new paper, researchers Abderaouf Bahi and Hasan Masrur introduce GridAnchor, a specialized verification layer designed to protect autonomous energy systems from "cyber-physical hallucinations" (CPHs).

TL;DR

GridAnchor is a physics-grounded verification layer that detects and repairs "hallucinations"—confident but physically incorrect outputs—from AI agents in autonomous power grids. By combining a Graph Transformer with statistical conformal prediction, it ensures that even if individual AI agents fail, the resulting grid actions remain safe and within physical limits.

What Problem Does This Paper Address?

As power grids become more complex, control is shifting from single controllers to multi-agent systems (MAS). In these chains, a Forecaster (F) sends data to a State Estimator (E), which informs a Dispatcher (D).

The authors identify a critical vulnerability: Cyber-Physical Hallucinations. Much like a Large Language Model (LLM) might confidently state a false fact, a power-system agent might emit a numeric claim that contradicts the laws of physics or sensor measurements. Because these agents are often "chained," a single hallucination can trigger a cascade, leading to unsafe physical actions like overvoltages or equipment damage.

Existing tools like "bad-data detection" focus on sensor errors, while "safety filters" only look at the final action. GridAnchor fills the gap by monitoring the entire communication chain, attributing errors to specific agents, and repairing the workflow in real-time.

Figure 1: A hallucination cascade (a) without and (b) with GridAnchor (schematic).

Methods

The researchers developed a multi-stage framework to "anchor" AI agents to physical reality:

  1. Taxonomy & Benchmark: They defined eight distinct types of CPHs (e.g., unit slips, stale replays, and physically impossible outputs) and created CPH-Bench to generate labeled training data.
  2. Physics-Residual Graph Transformer (PRGT): This is the core detection engine. It processes "residual tokens"—the difference between agent claims and what a linearized physical model of the grid (DistFlow) predicts. Because it uses a Graph Transformer, it inherently understands the grid's topology.
  3. Statistical Guardrails: To prevent "crying wolf," GridAnchor uses Mondrian conformal p-values. This provides a mathematically rigorous way to control the false-alarm rate, ensuring that detections are statistically significant.
  4. The Mitigation Loop: When a hallucination is detected, GridAnchor:
  5. Attributes the error to the root agent.
  6. Grounds the claim by replacing it with a verified reconstruction.
  7. Re-queries downstream agents with the corrected data.
  8. Repairs the final dispatch action using a Quadratic Program (QP) if the message is still flagged as risky.
Figure 2: Architecture of GridAnchor.

Datasets

The study utilized the SimBench dataset, a standard benchmark for electric power systems. Specifically, they tested on:

  • 1-MV-rural: A 99-bus rural medium-voltage feeder (primary testbed).
  • 1-MV-semiurb: A 122-bus semi-urban feeder (used to test model transferability). The authors also introduced CPH-Bench, a generator that injects the eight types of hallucinations defined in their taxonomy into these grid models.

Key Insights and Ideas

  • Physics as a Residual: You don't need a perfect, non-linear simulation to catch a hallucination. A linearized "physics-residual" is often enough to make a hallucination stand out from normal noise.
  • Cascades are the Real Threat: The paper reveals that while many agent errors are small, they compound. A hallucination in the Forecaster might seem minor but can lead the Dispatcher to take a catastrophic action.
  • Attribution is Key: Simply blocking a bad action isn't enough for an autonomous system. By attributing the error to a specific agent (e.g., "The Forecaster is hallucinating"), the system can "ground" that specific data point and allow the rest of the chain to continue functioning.
Table 2 Taxonomy of cyber-physical hallucinations.

Key Results

  • Detection Accuracy: GridAnchor achieved a message-level Area Under the Precision-Recall Curve (AUPRC) of 0.846, outperforming six strong baselines.
  • Grid Safety: In closed-loop simulations, GridAnchor resulted in zero voltage violations on test days. In contrast, an unguarded system left buses outside safe voltage bands in 0.07% of steps.
  • Operational Speed: The system is highly efficient, with a median latency of only 9.3 ms per 15-minute control step, making it practical for real-time deployment.
  • Robustness: The system remained effective even when 20% of meter readings were missing, proving its resilience to sensor failure.

Limitations and Future Work

While promising, the authors note several areas for improvement:

  • Simulation Bounds: The results are currently based on simulations. Real-world "field" hallucinations from deployed LLMs might be more subtle or complex than the synthetic types in CPH-Bench.
  • Model Dependency: GridAnchor relies on a reasonably accurate linearized model of the grid. If the grid topology changes significantly without the verifier being updated, performance could degrade.
  • Future Directions: The authors suggest moving toward "tolerance-aware" redundancy, where multiple agents perform the same task to provide a cross-check, further strengthening the "anchor."

For those interested in the implementation, project details and code are available on GitHub.

Summary updated

Abstract

Autonomous energy systems increasingly chain specialized learned agents, such as a forecaster, a state estimator and a dispatcher that may be a language model. These agents can emit confident outputs that contradict physics, and downstream agents may act on them. We define such cyber-physical hallucinations and their cascades, organize them in eight types, and build CPH-Bench, a generator of labeled episodes on SimBench feeders. We propose GridAnchor, a verification layer whose Physics-Residual Graph Transformer scores every claim from residuals between the claims, the measurements and a linearized grid model. Mondrian conformal p-values with the Benjamini-Hochberg procedure turn the scores into decisions, a root head attributes cascades, and a quadratic program repairs unsafe actions with a voltage margin calibrated at repaired steps. On a simulated rural medium-voltage feeder, GridAnchor reaches a message-level area under the precision-recall curve of 0.846, against 0.837 for the strongest of six baselines, a difference within the confidence interval. It leads on estimates and actions, whereas a self-consistency check that knows the label tolerances reaches 0.874 by leading on forecasts. The false-alarm rate stays within 0.17 percentage points of the nominal level on steps without injected hallucinations. In closed loop, GridAnchor leaves no bus outside the voltage band on the test days, against 0.0700% of the bus-steps without protection, and simpler detectors with the same mitigation come close. Stale but coherent messages and single-bus forecast errors remain mostly undetected.
BibTeX
@misc{bahi2026gridanchor,
  author = {Abderaouf Bahi and Hasan Masrur},
  title = {{GridAnchor: Specialized Multi-Agent AI for Safeguarding Autonomous Energy Systems Against Cyber-Physical Hallucinations}},
  year = {2026},
  eprint = {2611.25158},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.25158},
}
Figure