ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents
Authors: Arunkumar V, Gangadharan G. R., S. Senthilkumar
Organizations: University College of Engineering, BIT Campus, Anna University, Tiruchirappalli 620 024, India · Department of Computer Applications, National Institute of Technology Tiruchirappalli, Tiruchirappalli 620 015, India
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.
Figures & tables
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.
ρ
τT (cycles)
clean escalation
added cost/decision
covert FA
0.01
∞
0.000
$0.00
1.000
0.05
18.2
0.030
$1.33
0.868
0.10
15.2
0.064
$2.90
0.840
0.20
10.8
0.165
$7.44
0.775
Table 1: Escalation labour against attack reach on C-MAPSS FD001, clock k=5 , c_{\mathrm{esc}}=\45,all10,196testwindows.Ateach\rhothethresholdistheper−windowshapeatthat\rhowidenedbyProposition1(thelargerhalf’s\hat{\lambda}).Addedcostisrealisedescalationtimesc_{\mathrm{esc}}.Thelastcolumnisthepersistentcovertattacker’sfalse−alarmreach.At\rho=0.01the50calibrationenginesperhalfaretoofew(Proposition1needs99),so\hat{\lambda}=\infty$ and the monitor never abstains.
Figure 2: Why Vphys checks a decision rather than a value. An injected bulletin persuades the planner to request immediate maintenance on a unit whose clean interval is [42.1,97.9] cycles, while the call reports the remaining life truthfully. A value-level check sees 70∈I(x) and admits the call. The decision-level check derives the warranted class from the same interval, which is INSPECT, and rejects the mismatch.
Figure 3: Complementary coverage (Proposition 3 ). Each gate compares against a reference that one channel cannot move. The arms differ in kind: the Vphys cell holds by construction for every in-projection flip, while the Gconf cell holds only when the attacked prediction breaks the physics clock by more than τρ , and a slow persistent drift can move the anchors it is judged against (Section 9.7 ).
Clean ↓
Attack success rate ↓
Utility ↑
Configuration
err
abst
A1-a
A1-b
A2
A3
A1-a
Undefended ReAct
0.68
n/a
0.39
0.76
0.00
0.52
0.32
Defensive prompt only
0.41
n/a
0.29
0.18
0.54
0.67
0.48
Spotlighting
0.68
n/a
0.39
0.77
0.00
0.51
0.31
Adversarial training
0.41
n/a
0.29
0.18
0.04 †
n/a
0.48
Provenance monitor
0.41
0.00
0.29
0.00
0.54
0.67
0.48
Table 2: C-MAPSS FD001 with a Qwen3-14B planner: three-seed means over 100 queries per cell from the planner sweep, with the validator’s bounded retry. Undefended ReAct and Spotlighting use prompts without the decision rule; every other row uses the rule prompt, so each differs from Defensive prompt only by its monitor alone. Clean columns separate the two costs a defense imposes on clean traffic: err is the decision-error rate, the fraction of clean queries whose dispatched urgency class differs from πref(I(x)) , and abst is the false-abstention rate, the fraction escalated to a human. A2 and A3 count a harmful dispatch over the queries where one is possible ( y<25 or y>60 ). † Sensor-side value on all 10,196 test windows with the planner modelled by πref (Table 3 ); no planner was run behind these gates. An n/a marks a configuration with no abstention path, or an A3 cell that needs that planner run. Utility is task completion under A1-a. A1-b lies outside the protected projection by design.
Figure 4: Attack success rate by configuration on C-MAPSS FD001 with a Qwen3-14B planner, the cells of Table 2 . Outlined cells show a removed gate’s matched class returning: without Vphys , A1-a rises to 0.26; without Gconf , A2 rises to 0.88. A1-b lies outside the protected projection, and only composing a provenance monitor removes it.
Defence
clean escalation
fixed A2
adaptive
None
0.000
0.702
n/a
Adversarial training
n/a
0.037
0.485
Width-based conformal
0.050
0.235
0.925
Mahalanobis OOD
0.019
0.000
0.504
Physics clock ( k=30 )
0.027
0.006
0.046
Table 3: Sensor-channel defences on C-MAPSS FD001, all windows of the 100 test engines. The universal perturbation of Basak et al. [5] shifts predictions by 65.3 cycles on average and is a false-alarm attack: ungated, it flips 0.702 of healthy windows. The adaptive column is the same attacker re-optimised to stay under each defence’s own threshold, on 1,000 healthy windows for the learned defences and on every healthy window for the clock, whose one-shot covert attacker is exact.
Figure 5: C-MAPSS FD001 test engines, thresholds from Proposition 1 . (a) Covert reach on all 10,196 test windows (Table 4 ). A fixed physical range is blind to suppression and leaves a covert attacker reaching 0.529 of healthy windows; anchoring to the asset’s own history cuts false-alarm reach to 0.046 at k=30 , and at k=10 and k=30 closes suppression. (b) Fixed and adaptive attacks (Table 3 ). A Mahalanobis detector, adversarial training and the clock all stop the fixed universal perturbation; the first two fall to an attacker who re-optimises under their own threshold, measured on 1,000 healthy windows for the learned defences and on every healthy window for the clock. No adaptive attack applies to the undefended pipeline. Only the physics clock carries a bound into the adaptive setting (Proposition 2 ).
Monitor
clean escalation
suppression
false alarm
λ^
Static support
0.004
0.151
0.529
1.02, 1.02
Clock, k=5
0.034
0.005
0.033
1.44, 1.40
Clock, k=10
0.030
0.000
0.049
1.18, 1.30
Clock, k=30
0.027
0.000
0.046
1.00, 1.00
Table 4: Covert reachability on all 10,196 windows of the 100 C-MAPSS FD001 test engines. Each threshold is the per-window shape τ0 from the ten held-out validation engines, widened by the λ^ of Proposition 1 , cross-fit over two halves of the test engines (last column, one value per half). q^=27.89 cycles, ρ=0.05 , attacker budget 65 cycles. Suppression is reachable when the attacker can raise a critical asset ( y<25 ) into no_action ; false alarm when it can lower a healthy asset ( y>60 ) into maintain_now .
Per-window
Proposition 1
Data
Monitor
pooled
mean
p90
λ^
pooled
mean
p90
FD001, 100 engines
static support
0.058
0.058
0.124
1.02, 1.02
0.004
0.004
0.000
clock, k=5
0.140
0.115
0.334
1.44, 1.40
0.034
0.030
0.084
clock, k=10
0.077
0.065
0.194
1.18, 1.30
0.030
0.027
0.092
clock, k=30
0.027
0.023
0.076
1.00, 1.00
0.027
0.023
0.076
FD004, 248 engines
static support
0.061
0.059
0.196
1.02, 1.02
0.000
0.000
0.000
Table 5: Clean false abstention of the deployed monitor (a window enters the history only if it passes) on every test window, at ρ=0.05 . Per-window : thresholds from per-window quantiles on the validation engines. Proposition 1 : the same thresholds widened by λ^ , chosen on one half of the test engines and applied to the other (one value per half). Pooled is the share of all clean windows escalated; mean and p90 are taken over assets, each asset’s own share, so mean is the quantity the proposition bounds.
Per-window
Proposition 1
Covert reach
Pattern
Observed
mean
worst
λ^
mean
worst
per-window
Prop. 1
Smooth
truncated
0.046
0.072
1.09
0.034
0.055
0.049
0.055
Noisy
truncated
0.031
0.040
1.03
0.027
0.036
0.266
0.285
Abrupt onset
truncated
0.023
0.045
1.04
0.021
0.045
0.631
0.659
Switching regimes
truncated
0.038
0.074
1.08
0.027
0.046
0.049
0.055
Rate mismatch
truncated
0.030
0.041
1.03
0.026
0.041
0.036
0.038
Table 6: Simulated fleets, monitor at k=30 , ρ=0.05 , 10 repetitions of 400 test trajectories each. Mean is the per-asset false-abstention share averaged over assets and repetitions, and worst the largest per-repetition mean. Covert reach is the exact one-shot false-alarm reach on healthy windows (budget 65 cycles), the detection price of widening the thresholds. The last row calibrates on one pattern and deploys on another, which breaks the assumption of Proposition 1 .
Attack
Monitor
caught before harm
FA success
Universal perturbation
static support
0.000
1.000
clock, k=30
1.000
0.142
Adaptive, non-covert
static support
0.000
1.000
clock, k=30
1.000
0.138
Adaptive, covert, persistent
static support
0.000
1.000
clock, k=30
0.000
0.638
Table 7: Sensor attacks along full trajectories on all 100 FD001 test engines, attacker persistent from the degradation onset. Caught is the fraction of attacked assets escalated before any harmful dispatch; FA is the false-alarm success rate over healthy windows. The covert attacker is idealised: it may place the prediction anywhere in the ϵ -ball’s image and re-plans every window, an upper bound on attack power.
Suppression ↓
False alarm ↓
Dataset
clean abst
static
clock
static
clock
C-MAPSS FD001
0.027
0.151
0.000
0.529
0.046
C-MAPSS FD004
0.030
0.808
0.099
0.221
0.038
MetroPT-3
0.075
0.500
0.183
0.998
0.104
Table 8: One-shot covert reach by dataset, static support against the physics clock, at the clean escalation rate each clock pays. All cells are sensor-side and need no planner; the planner sweep is FD001 only (Table 2 ). C-MAPSS rows score every test window, with thresholds from Proposition 1 cross-fit over halves of the test engines. MetroPT-3 has two episodes, too few for the proposition, and keeps per-window thresholds cross-fit across them. The horizon is k=30 windows on C-MAPSS and k=10 on MetroPT-3, whose ten-minute windows make thirty a five-hour anchor.
ρπ , clean
Vphys trig.
Clean abst
A1-a ASR
Planner
all
worst
under A2
after retry
prompt
+Vphys
Qwen3-14B
0.410
0.723
0.340
0.000
0.287
0.000
Phi-4-14B
0.500
1.000
0.393
0.000
0.013
0.000
Gemma-3-12B
0.627
0.864
0.820
0.000
0.117
0.000
Mistral-Nemo-12B
0.648
1.000
0.703
0.000
0.090
0.000
Llama-3.1-8B
0.456
0.907
0.360
0.000
0.310
0.000
Table 9: Planner adherence on FD001 under the rule prompt, five open planners, three seeds of 100 queries per attack. ρπ is the first-answer adherence failure on clean queries and worst the same rate in the warranted class where it is highest; the next column is Vphys ’s trigger rate under A2, which Proposition 3 (b) bounds by the worst-class rate. After retry is the clean escalation rate of Vphys once the bounded retry has run.
Artificial intelligence agent security studies attacks and safeguards for tool-using, persistent, and collaborative systems across digital and physical settings.