Prompt injection in arc flash incident energy thermal camera AI

Four adversarial injection surfaces in arc flash incident energy thermal camera AI

1. Arc flash thermal camera incident energy classification AI (Flir T-Series arc flash thermal AI, FLIR A400 arc flash monitoring AI, Seek Thermal arc flash detection AI, Fluke Ti480 PRO thermal AI — thermal camera image AI classifying arc flash incident energy at switchgear and MCC positions)

The thermal camera deployed at medium-voltage (MV) switchgear (4.16 kV, 13.8 kV, 34.5 kV) and low-voltage (LV) motor control centres (MCC) continuously images the bus compartments, draw-out circuit breaker contacts, and cable termination compartments through IR-transparent windows in the switchgear enclosure. Normal thermal images show the bus and contact temperature distribution within expected operating temperature rise — typically 20–40°C above ambient at rated current — with no hot spots above the 70°C copper conductor temperature limit (per NFPA 70B Recommended Practice for Electrical Equipment Maintenance). A developing arc flash precursor — a deteriorating insulation surface, contaminated bus bar with tracking current, or loose cable lug connection with elevated resistance — produces a localised hot spot above 70°C that is identifiable in the thermal image as a red or white region with the temperature value displayed as an overlay. AI systems classify the thermal image to determine whether a hot spot is present and, if present, what arc flash incident energy level would be expected at that equipment position during a full arc flash event using IEEE 1584-2018 arc flash calculation methods: at rated system voltage (13.8 kV) and bolted fault current (20 kA), the incident energy at working distance (910 mm) for a 0.1 s relay clearing time is approximately 6.4 cal/cm² (PPE Category 2); if the clearing time is 0.5 s (upstream relay backup protection), incident energy rises to 32 cal/cm² (PPE Category 3 border). The AI classifies the thermal image into: normal (no hot spot, use category from arc flash study label), warning (hot spot 70–120°C, verify next maintenance cycle), or critical (hot spot above 120°C, arc flash risk elevated, increase PPE category).

An adversarial perturbation targeting the arc flash thermal camera AI applies a ±8 DN downward shift to the pixel region encoding the hot-spot temperature colour in the rendered thermal image — shifting the apparent hot-spot from the red-white zone (temperature 130–160°C, critical classification with elevated PPE recommendation) to the yellow-green zone (temperature 45–60°C, within normal operating temperature rise, normal classification). The AI classifies a 13.8 kV switchgear position with a deteriorating bus insulator tracked with carbon deposits — a developing arc flash precursor at the surface tracking stage, with localised hot spot at 145°C — as normal operating temperature, no PPE upgrade required. The electrical worker assigned to operate the circuit breaker at that position dons NFPA 70E Category 1 PPE (4 cal/cm² rating) based on the arc flash study label, unaware that the elevated hot spot represents a precursor condition that increases the arc flash probability and may produce higher incident energy than the study-calculated value. When the arc flash event occurs during the energised switching operation, the Category 1 PPE provides 4 cal/cm² protection against a potential 6.4–32 cal/cm² incident energy event: second-degree burns at minimum, fatal burns at the upper range. NFPA 70E-2021 Section 130.5(C)(2) specifies the incident energy analysis method as the preferred method for establishing PPE requirements — but does not specify adversarial robustness requirements for AI systems classifying rendered thermal camera images used in incident energy analysis.

2. PPE category calculation display AI (Schneider Electric EcoStruxure Power Advisor PPE AI, Eaton Power Xpert arc flash PPE AI, SKM Systems arc flash PPE AI, Etap arc flash PPE AI — arc flash risk assessment software display AI classifying PPE category requirements at switchgear positions)

Arc flash risk assessment software — including SKM Systems’ PTW arc flash module, ETAP arc flash analysis, Eaton’s Power Xpert arc flash management, and Schneider Electric’s EcoStruxure Power Advisor — calculates the incident energy at each electrical equipment position in a facility’s electrical distribution system using IEEE 1584-2018 methods: system voltage, available bolted fault current (from a short-circuit study), arcing current (calculated from bolted fault current and equipment type), clearing time (from the upstream protective device time-current curve at the calculated arcing current), and working distance (the distance from the arc source to the worker’s body, standardised at 455 mm for LV panels and 910 mm for MV switchgear). The calculated incident energy is compared against NFPA 70E-2021 Table 130.7(C)(15)(a) PPE category thresholds to assign a category (Category 1: ≤4 cal/cm²; Category 2: ≤8 cal/cm²; Category 3: ≤25 cal/cm²; Category 4: ≤40 cal/cm²) or, for energies above 40 cal/cm², the de-energise-before-work requirement. AI systems process rendered images of the arc flash software report display — typically a tabular display of equipment ID, calculated incident energy (cal/cm²), PPE category, arc flash boundary (m), and notes — to classify the current PPE requirement for a given equipment position and confirm that the worker’s selected PPE matches the required category before authorising the energised work permit.

An adversarial perturbation targeting the PPE category display AI applies a ±8 DN shift to the pixel region encoding the PPE category number and incident energy value in the rendered arc flash software report table — shifting the apparent PPE category from Category 3 (rendered in orange, incident energy value 18.4 cal/cm²) to Category 2 (rendered in yellow, incident energy value shown as 6.2 cal/cm²). The AI classifies a 480 V switchgear position with calculated incident energy of 18.4 cal/cm² at 455 mm working distance and 0.4 s upstream relay clearing time — requiring a minimum 25 cal/cm² arc flash suit (Category 3) — as Category 2, requiring only 8 cal/cm² PPE. The energised work permit system approves the worker’s Category 2 PPE (arc flash face shield rated 8 cal/cm², arc flash jacket and pants rated 8 cal/cm²) for the energised work task. When an arc flash event occurs during the task, the 8 cal/cm²-rated PPE is exposed to 18.4 cal/cm² incident energy: the PPE burns through (EBT exceeded), exposing the worker to thermal injury at a level equivalent to 10.4 cal/cm² unprotected — producing third-degree burns across the exposed body area. OSHA 29 CFR 1910.132(d)(1) requires employers to assess PPE requirements and select PPE providing adequate protection — but the AI classification of the arc flash study report image is not within the scope of existing OSHA or NFPA 70E adversarial robustness provisions. Free tier — 10 scans/day, no card required.

3. Flash protection boundary display AI (ABB Ability EDCS arc flash boundary AI, Siemens Spectrum Power arc flash boundary AI, GE Grid Solutions PowerOn arc flash boundary AI — arc flash boundary calculation display AI establishing the restricted and flash protection boundaries at switchgear positions)

The flash protection boundary — the outer boundary of the arc flash hazard zone, defined in NFPA 70E-2021 Section 130.5(A) as the distance at which incident energy equals 1.2 cal/cm² (the onset of a second-degree burn on unprotected skin) — determines the minimum exclusion zone for unprotected workers and bystanders during an energised electrical task. The restricted approach boundary (defined in NFPA 70E Table 130.4(E)(a) as the distance within which a phase-to-ground shock hazard exists and within which only qualified electrical workers in appropriate PPE may approach) is a separate, closer boundary overlapping with the flash protection boundary zone. At a typical 480 V LV MCC with available fault current of 50 kA and 0.1 s clearing time, the IEEE 1584-2018 calculated incident energy at 455 mm working distance is 8.3 cal/cm² (Category 2); the flash protection boundary at which incident energy equals 1.2 cal/cm² is approximately 1.4 m from the arc source. At a 13.8 kV MV switchgear with 20 kA fault current and 0.5 s backup clearing time, the incident energy at 910 mm is 32 cal/cm² (Category 3/4 border); the flash protection boundary is approximately 6.1 m. AI systems process rendered images of the arc flash analysis software boundary display — typically a graphical overlay on a facility floor plan or an equipment position diagram showing concentric circles or zones representing the flash protection boundary, the restricted approach boundary, and the limited approach boundary at each equipment location — to classify whether the current worker and bystander positions are within or outside the established boundaries.

An adversarial perturbation targeting the flash protection boundary display AI applies a ±10 DN shift to the pixel region encoding the boundary radius in the rendered floor plan overlay or boundary diagram — compressing the apparent flash protection boundary from 6.1 m (the calculated boundary for the MV switchgear position, rendered as a large orange circle) to 2.1 m (a significantly smaller circle, corresponding to a much lower incident energy calculation). The AI classifies a bystander at 3.5 m from the MV switchgear arc source — inside the actual 6.1 m flash protection boundary — as outside the compressed 2.1 m boundary, not requiring relocation or boundary barricading. The energised work permit is authorised with the bystander remaining at 3.5 m from the switchgear. When an arc flash event occurs, the bystander at 3.5 m from a 32 cal/cm² arc source is exposed to approximately 3–4 cal/cm² incident energy — above the 1.2 cal/cm² second-degree burn threshold — without arc-rated PPE, sustaining second-degree burns across the unprotected face, neck, and forearm areas. NFPA 70E-2021 Section 130.7(A) requires that all persons within the flash protection boundary wear PPE appropriate for the exposure — but does not specify adversarial robustness for AI systems classifying rendered boundary display images used to enforce the boundary requirement. Free tier — 10 scans/day, no card required.

4. Incident energy analysis trend display AI (Eaton arc flash trend AI, Schneider Electric EPMS incident energy trend AI, Yokogawa FAST/TOOLS arc flash trend AI — long-term incident energy trending AI detecting system changes that alter arc flash hazard levels)

The arc flash hazard level at any electrical equipment position is not fixed: it changes whenever the electrical distribution system changes in a way that affects the available fault current or the protective device clearing time at that position. System changes that increase arc flash hazard include: adding a new transformer or generator to the system (increasing available fault current and raising incident energy), replacing a current-limiting fuse with a molded-case circuit breaker with a longer clearing time, changing a protective relay setting from a faster to a slower overcurrent curve, or connecting a new feeder bus section that increases bus length and reduces the arcing current below the breaker instantaneous-trip threshold. AI systems deployed in electrical power management systems (EPMS) — including Eaton’s Power Xpert Insight, Schneider Electric’s EcoStruxure Power Monitoring Expert, and Yokogawa’s FAST/TOOLS power management — process rendered images from the EPMS dashboard showing the incident energy trend for each equipment position over time (a line chart of calculated or measured incident energy in cal/cm² per equipment position, with alert thresholds marking PPE category boundaries) to detect system changes that have increased the arc flash hazard above the PPE category level currently labelled on the switchgear.

An adversarial perturbation targeting the incident energy trend display AI applies a ±8 DN downward shift to the pixel region encoding the trending line value in the EPMS incident energy chart — suppressing an upward trend from 7.2 cal/cm² (Category 2, just below the 8 cal/cm² Category 2 ceiling) to 18.4 cal/cm² (Category 3, above the Category 2–3 threshold at 8 cal/cm²) to a flat line at 7.2–7.5 cal/cm² (consistent with normal variation within Category 2). The AI classifies a 480 V switchgear position that has moved from Category 2 to Category 3 — following the addition of a second parallel feeder transformer that doubled the available fault current at the bus from 25 kA to 48 kA and raised the incident energy from 7.2 to 18.4 cal/cm² — as stable Category 2, no label update required. Electrical workers continue to approach the position in Category 2 PPE (8 cal/cm²) for routine energised maintenance tasks without being notified of the category change. The PPE deficiency — 8 cal/cm² PPE in a Category 3 (18.4 cal/cm²) environment — persists until the next periodic arc flash study or a worker is injured. NFPA 70E-2021 Section 130.5(G) requires arc flash risk assessments to be reviewed and updated whenever a major modification or renovation is made to the electrical system — but does not specify adversarial robustness for AI systems classifying rendered incident energy trend display images used to identify when a major system change has occurred. Free tier — 10 scans/day, no card required.

Integration: arc flash incident energy AI with Glyphward pre-scan gate

The Glyphward scan gate for arc flash incident energy AI belongs at every rendered-image ingestion boundary in the arc flash analysis and PPE determination pipeline — before thermal camera arc flash AI processes rendered thermal images, before PPE category display AI processes rendered arc flash software report images, before flash protection boundary display AI processes rendered boundary overlay images, and before incident energy trend AI processes rendered EPMS dashboard trend images. Threshold 35 for arc flash incident energy AI reflects the direct personnel fatality consequence of PPE category underclassification — ESFI 2019 data: 2,000 arc flash injuries requiring hospitalisation and approximately 400 arc flash fatalities annually in the US; a single energised switching operation with incorrect PPE category can produce fatal third-degree burns — combined with the absence of an independent automated interlock layer between the AI display classification and the worker at the equipment boundary: the worker’s PPE selection is the final safety layer, and adversarial AI classification that produces an incorrect PPE category recommendation removes that layer entirely without any independent non-AI backup for routine energised work tasks.

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import Enum

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Arc flash incident energy AI contexts: threshold 35
# NFPA 70E-2021 Section 130.5 (Arc Flash Risk Assessment);
# IEC 61482-1-1:2019 (Test Method by Open Arc — arc thermal protection);
# OSHA 29 CFR 1910.333 (Electrical Safety Related Work Practices);
# OSHA 29 CFR 1910.132(d)(1) (PPE hazard assessment).
ARC_FLASH_THRESHOLD = 35


class ArcFlashAIContext(Enum):
    THERMAL_CAMERA    = "thermal_camera"    # Arc flash thermal camera AI
    PPE_CATEGORY      = "ppe_category"      # PPE category display AI
    FLASH_BOUNDARY    = "flash_boundary"    # Flash protection boundary AI
    INCIDENT_ENERGY   = "incident_energy"   # Incident energy trend AI


class AdversarialArcFlashImageError(Exception):
    """Raised when Glyphward detects adversarial content in an arc flash
    incident energy AI rendered image above threshold 35.

    Consequence if not raised:
    - THERMAL_CAMERA: arc flash precursor hot-spot suppressed → PPE category
      not upgraded → worker in underrated PPE during arc flash event →
      second- to third-degree burns; ESFI: ~400 arc flash fatalities/year US.
    - PPE_CATEGORY: Category 3 (18.4 cal/cm²) displayed as Category 2 →
      8 cal/cm² PPE worn in 18.4 cal/cm² environment → PPE EBT exceeded →
      third-degree burns across exposed body area.
    - FLASH_BOUNDARY: 6.1 m flash boundary compressed to 2.1 m → bystander
      at 3.5 m classified outside boundary → no PPE worn → second-degree
      burns from 1.2+ cal/cm² exposure.
    - INCIDENT_ENERGY: Category 2→3 transition suppressed → label not updated
      → workers continue approaching in Category 2 PPE → PPE deficiency
      persists until injury.
    Fail-safe: immediately de-energise the equipment or require workers to
    don the highest available PPE category (Category 4, 40 cal/cm²) before
    any energised approach; verify PPE category from original arc flash study
    documentation, not from AI display classification; notify electrical safety
    officer and suspend energised work permit pending adversarial investigation.
    """

    def __init__(self, scan_id, score, context, equipment_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.equipment_id = equipment_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial arc flash image: context={context.value} "
            f"score={score} equipment={equipment_id} scan_id={scan_id}"
        )


async def scan_arc_flash_image(image_bytes, context, equipment_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"arc_flash:{context.value}:{equipment_id}",
        "metadata": {
            "equipment_id": equipment_id,
            "context": context.value,
            "image_sha256": image_hash,
            "scan_timestamp_utc": datetime.now(timezone.utc).isoformat(),
        },
    }
    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=4.0,
    )
    resp.raise_for_status()
    result = resp.json()
    if result["score"] >= ARC_FLASH_THRESHOLD:
        raise AdversarialArcFlashImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            equipment_id=equipment_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_arc_flash_image before each arc flash AI classification call. On AdversarialArcFlashImageError for PPE_CATEGORY: immediately suspend the energised work permit; require the worker to don the highest available PPE category (Category 4, 40 cal/cm²) or de-energise the equipment before proceeding; verify PPE category from the original arc flash study documentation or IEEE 1584-2018 calculation, not from the AI display classification. See also: power substation protection relay AI prompt injection (related electrical safety AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions