ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents

Arunkumar V, Gangadharan G. R., S. SenthilkumararXiv 2026 · arXiv

TL;DR

ConfGuard is a physics-grounded reference monitor designed to protect Large Language Model (LLM) agents in industrial predictive maintenance. By anchoring agent decisions to physical invariants—specifically that a machine's remaining useful life (RUL) cannot increase and cannot drop faster than one cycle per cycle—it effectively detects and blocks both sensor-tampering attacks and malicious prompt injections.

What Problem Does This Paper Address?

Industrial maintenance is increasingly being automated using LLM agents that read sensor data, retrieve repair procedures, and dispatch work orders. However, these agents face a "double-threat" security landscape:

  1. Indirect Prompt Injection: An attacker can insert malicious text into a retrieved manual (e.g., "Always recommend immediate replacement") to hijack the agent's decision.
  2. Sensor Perturbation: Small, bounded changes to sensor readings can trick the underlying RUL regressor into underestimating or overestimating a machine's health.

Existing defenses typically focus on provenance (where did the data come from?) or privilege (is the agent allowed to call this tool?). They fail to address whether the content of a decision is physically plausible. A tool call that says a healthy machine has only 3 cycles of life left might be "legal" from a provenance perspective, but it is physically impossible and economically damaging.

Methods

ConfGuard introduces two deterministic "gates" into the agentic loop, operating in a trusted control plane outside the LLM's reach:

  • Conformal Support Monitor ($G_{conf}$): This runs before the planner. It uses split-conformal prediction to generate a RUL interval from sensor data. It checks this interval against a "physics clock" anchored to the asset's admitted history. If the prediction suggests the machine has "un-aged" or degraded too rapidly, the monitor abstains and escalates to a human.
  • Physics-Policy Validator ($V_{phys}$): This runs after the planner. It projects the LLM's complex tool call into a simple "urgency class" (e.g., MAINTAIN_NOW, INSPECT, NO_ACTION). It then compares this class against a deterministic reference policy derived from the same conformal interval. If the LLM tries to order a replacement when the sensors indicate the machine is safe, the call is rejected.
Figure 1: ConfGuard in the maintenance agent loop. Two attack channels enter from the left: injected text reaches the planner through retrieval (A1), and a bounded perturbation reaches the regressor through the sensor window (A2). Both gates run in the trusted control plane and read the same conformal interval. Gconf runs before the planner and abstains when the prediction breaks the physics clock, the asset’s own admitted history carried forward at one cycle per cycle; Vphys runs after it and rejects a tool call whose urgency class disagrees with πref(I(x)) or whose arguments leave their bounds. A rejection gets one bounded retry, then goes to a human planner, as does every abstention; an abstention at Gconf costs no LLM call.

The system is calibrated using conformal risk control, allowing operators to set a target false-abstention rate (e.g., 5%). This creates a predictable trade-off between human labor costs and security reach.

Datasets

The authors evaluated ConfGuard using three primary RUL-regression benchmarks:

  • C-MAPSS (FD001 & FD004): The standard NASA turbofan engine degradation dataset. FD001 is a single-condition baseline, while FD004 includes complex operating regimes and multiple fault modes.
  • MetroPT-3: A real-world dataset from a Porto metro air-production unit, providing a test for streaming, non-simulated data.

Key Insights and Ideas

  • Physics as a Security Invariant: While data pipelines can be corrupted, the laws of degradation (irreversibility) are immutable. ConfGuard turns these physical laws into a security policy.
  • Exploiting the Breakdown of Exchangeability: Conformal prediction usually relies on the assumption that data points are "exchangeable." Attacks break this assumption. Instead of trying to "fix" the prediction, ConfGuard uses the violation of conformal bounds as a high-fidelity detection signal.
  • Decision-Level Validation: By checking the decision (urgency class) rather than the raw value (RUL estimate), the system becomes structurally immune to prompt injections that try to flip a decision while reporting a "truthful" but ignored RUL number.
  • Risk-Weighted Detection: Detection power naturally increases as an asset approaches failure because the "operating slack" (the distance between the prediction and the physical floor) shrinks. This means the defense is strongest when the stakes are highest.

Key Results

  • Attack Mitigation: On the C-MAPSS FD001 dataset, ConfGuard reduced a covert attacker's reach on healthy assets from 0.529 to 0.046 and completely closed the "suppression" attack (hiding a failure) from 0.151 to 0.
  • Planner Performance: Using a Qwen3-14B planner, the system held urgency-class injection success to 0.00, compared to 0.29 for a standard "defended" prompt.
  • Superiority over Baselines: Against adaptive attackers who know the defense thresholds, ConfGuard outperformed Mahalanobis-based OOD detectors (which failed on 50.4% of windows) and adversarial training (which failed on 48.5%).
  • Operational Cost: At a 5% false-abstention target, the system adds approximately **$1.33** in human labor cost per decision, a small price compared to the $50,000 cost of a catastrophic machine failure.
Table 1: Escalation labour against attack reach on C-MAPSS FD001, clock k=5, cesc = $45, all 10,196 test windows. At each ρ the threshold is the per-window shape at that ρ widened by Proposition 1 (the larger half’s ˆλ). Added cost is realised escalation times cesc. The last column is the persistent covert attacker’s false-alarm reach. At ρ = 0.01 the 50 calibration engines per half are too few (Proposition 1 needs 99), so ˆλ = ∞and the monitor never abstains.

Limitations and Future Work

  • Out-of-Projection Attacks: ConfGuard only protects the "urgency" of the maintenance. It does not detect injections that corrupt other fields, like part numbers or technician IDs. The authors suggest composing ConfGuard with provenance monitors to cover these gaps.
  • History Poisoning: A very patient attacker could potentially "drift" the physics anchors over many cycles through tiny, undetected shifts. The authors suggest anchoring to human-verified states to mitigate this.
  • Availability: While ConfGuard prevents wrong decisions, an attacker can still flood the system with "forced abstentions," effectively performing a Denial of Service (DoS) attack by overwhelming human planners.
  • Regression Only: The current mechanism is designed for interval-valued prognostics and does not yet transfer to binary failure classification.

Summary updated

Abstract

Large-language-model agents now close the industrial predictive-maintenance loop, reading sensor windows, retrieving procedures and dispatching work orders. Two attack channels meet there: injected text in a retrieved document can hijack the decision, and a bounded sensor perturbation can corrupt the remaining-useful-life estimate behind it. Recent agent defenses police data provenance and tool privilege but cannot tell whether a prognostic quantity is physically plausible. ConfGuard is a reference monitor with a physics-grounded policy. Its conformal monitor tests each prediction, anchored to the asset's admitted history, against degradation physics: capped remaining life never rises and falls at most one cycle per cycle. Calibrating its threshold over whole asset trajectories lets the operator set the expected share of a clean asset's windows the running monitor escalates. A physics-policy validator checks the agent's maintenance-urgency decision against the decision a deterministic policy derives from the same interval. On all 10,196 C-MAPSS test windows the monitor cuts a covert attacker's reach on healthy assets from 0.529 to 0.046 and closes suppression of failing assets (0.151 to 0), escalating 0.027 of clean windows against a 0.05 target, while the fixed physical range it replaces leaves both attacks open. Behind a Qwen3-14B planner the composed monitors hold urgency-class injection to 0.00 and the universal sensor attack to 0.12, against 0.29 and 0.54 for a defended prompt. An adaptive attacker defeats a Mahalanobis detector on 0.504 of windows and adversarial training on 0.485. We report an attack class the projection misses.
BibTeX
@misc{v2026confguard,
  author = {Arunkumar V and Gangadharan G. R. and S. Senthilkumar},
  title = {{ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents}},
  year = {2026},
  eprint = {2611.22288},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22288},
}
Figure