Prompt injection in steel electric arc furnace EAF AI

Four adversarial injection surfaces in steel electric arc furnace EAF AI

1. Furnace shell water-cooled panel thermal camera AI (FLIR A615 thermal AI, Cognex In-Sight thermal AI, Teledyne FLIR thermography AI — EAF furnace shell WCP hotspot classification AI)

Modern EAF furnace shells — the cylindrical steel structure forming the sidewall and roof of the arc furnace above the refractory slag line — are constructed with water-cooled panels (WCPs) as their principal heat-removal elements. A WCP is a fabricated copper or steel panel with internal water cooling channels, forming a section of the furnace sidewall or roof and positioned above the refractory brick lining that contacts the molten steel and slag bath. WCPs in a typical 150-tonne EAF shell number 20–40 individual panels, each with dimensions of approximately 1.5 m × 1.5 m × 80–120 mm wall thickness. Cooling water circulates through the internal channels at flow rates of 15–40 m³/h per panel, with inlet temperature of approximately 45–60°C and outlet temperature of 65–85°C (maximum), removing approximately 30–40% of the total electrical energy input to the furnace shell (the remainder exits via the off-gas system and the steel/slag tap). The internal surface temperature of a correctly functioning WCP — the face exposed to the arc radiation and refractory gas above the slag line — operates at 80–200°C in normal furnace operation, depending on arc position and radiation geometry. When an arc attachment event, refractory failure, or process deviation causes a localised concentration of arc radiation or conductive heat flux on a single panel — from an electrode operating closer to the sidewall than design, a refractory brick dropout creating a direct sightline between the arc and the WCP surface, or a slag splashing event that coats a WCP with molten slag and increases thermal conductivity — the external surface temperature of the affected WCP rises above the normal range. At external surface temperatures above approximately 200–280°C, the internal surface of the panel (steel nominal thickness 8–12 mm for the innermost hot-face layer in a fabricated steel WCP) approaches the onset of localised melting in any thin-wall or defect region, and the wall can burn through — creating a breach in the WCP through which pressurised cooling water (at 4–8 bar) discharges directly into the EAF interior and contacts the molten steel bath at 1,560–1,700°C.

The physical consequence of water contacting molten steel is determined by the fundamental thermodynamics of explosive steam generation: water at 20°C contacting a steel surface at 1,600°C vaporises in a time on the order of microseconds, generating steam at the boiling point (100°C at 1 bar) with a specific volume approximately 1,680 times that of liquid water at the same conditions — but at the actual furnace temperature and the confined geometry of the WCP breach point, the effective volume expansion factor in the initial milliseconds of contact approaches 5,000:1. The resulting pressure pulse — a steam explosion — carries an instantaneous energy release equivalent to 30–50 kg of TNT for a moderately sized WCP breach allowing 5–10 litres of water per second into the bath, sufficient to rupture the furnace shell, project molten steel and refractory fragments through the EAF building, destroy the electrode support structure, and create an arc flash from the disrupted electrode circuit. EAF thermal camera AI systems process rendered false-colour thermal images of the furnace shell exterior (a thermal colour map with a temperature scale: blue < 100°C normal, yellow 100–200°C elevated, red 200–280°C critical, white > 280°C emergency burnthrough risk) generated by FLIR A615 radiometric thermal cameras, Cognex In-Sight thermal vision systems, or Teledyne FLIR industrial thermography systems mounted at fixed positions around the EAF shell exterior to classify WCP condition: normal, elevated, critical, and emergency-burnthrough-risk. The AI classification drives cooling water flow increase (opening the panel flow control valve), arc power reduction (reducing the electrode current to reduce radiated thermal flux on the affected panel), or emergency shutdown (opening the main arc circuit breaker and initiating emergency scram).

An adversarial perturbation targeting the EAF furnace shell thermal camera AI applies a ±8 DN colour shift in the pixel region encoding a developing WCP thermal hotspot — shifting the apparent hotspot colour from the red or white emergency range to the yellow elevated range in the rendered thermal false-colour image. The AI classifies a developing WCP burnthrough precursor (panel exterior temperature 220–280°C, indicating interior surface approaching the water boiling point and panel wall beginning to thin from localised melting) as an elevated-but-monitoring condition rather than a critical alarm requiring immediate intervention. Cooling water flow is not increased, arc power is not reduced, and the panel is not scheduled for replacement at the next tap. The WCP exterior temperature continues rising as the arc radiation continues to impinge: the panel wall reaches the melting point at the hotspot location (steel melting point approximately 1,370–1,510°C depending on carbon content) and a breach opens. The SSAB Oxelösund EAF WCP burnthrough incident in 2015 — in which a water-cooled panel on the EAF furnace shell burned through during a heat, resulting in steam explosion and arc flash that injured four workers — established the documented consequence of undetected WCP thermal overloading in an operating furnace. Nucor Corporation EAF water-cooled panel failures at multiple US mini-mill sites (early 2000s), resulting in steel splash onto primary cooling circuits and OSHA citations for inadequate WCP inspection protocols, established the regulatory enforcement baseline. AIST T-13 (EAF Safety and Health Standards, formerly AISE 12.1) requires periodic WCP inspection and maintenance based on thermocouple and visual inspection criteria — but does not address adversarial robustness requirements for AI systems classifying the rendered thermal camera images that automate the continuous monitoring between manual inspection intervals.

2. Wet scrap charge bucket detection camera AI (OSHA 29 CFR 1910.179 scrap inspection AI, Danieli EAF scrap bucket camera AI, SMS group Quantum scrap quality AI — wet scrap moisture detection camera classification AI)

Electric arc furnace steelmaking uses scrap steel — obsolete steel from demolition, post-consumer vehicles, industrial offcuts, and internal home scrap — as its primary charge material, charged into the EAF in large cylindrical buckets with a clam-shell or drop-bottom design (capacity 80–120 tonnes per bucket for a 150-tonne EAF heat) via overhead charging cranes. OSHA 29 CFR 1910.179 (Overhead and gantry cranes) governs the design, inspection, and operation of the overhead crane systems used to lift and position scrap charge buckets above the open EAF roof during charging — including requirements for load-rating markings, annual inspection, and pre-lift inspection of the load. AIST T-13 Section 4.2 (Scrap Preparation) establishes the steel industry’s safety requirements for scrap charging, explicitly prohibiting the charging of wet or ice-contaminated scrap and requiring visual inspection of the scrap charge for moisture before charging. The hazard is well-established and documented in the steel industry: scrap steel arriving at EAF facilities from scrap yards may contain trapped water in hollow sections (pipes, I-beams, closed automotive body sections), pooled water at the bottom of the charge bucket from precipitation during outdoor storage or rain during transport, ice formed in cavities during winter operations, or water-bearing contamination from hydraulic fluid and cutting coolant on machine scrap. When a wet scrap charge is dumped from the bucket through the open furnace roof into the EAF and the scrap cascades into the molten steel bath at 1,560–1,700°C, the water — whether free-pooled water at the bucket bottom, water trapped in hollow scrap sections, or ice in protected cavities — contacts the molten bath in a time of less than 1 second and vaporises explosively.

The physics of wet scrap steam explosion are severe: even 1–2 litres of water trapped in a 100-tonne scrap charge can produce a steam explosion equivalent to 10–20 kg of TNT (instantaneous energy release in < 1 ms, water hammer impulse), generating a pressure pulse that ruptures the furnace shell, projects molten steel fragments and refractory shrapnel upward through the furnace roof opening and outward through furnace shell gaps into the surrounding melt shop. The furnace roof and electrode arms are directly above the burst zone, and the melt shop operating floor is typically 5–15 metres from the furnace shell. Steel industry statistical data compiled by AIST indicates that wet scrap steam explosions account for approximately 40–60% of all EAF fatalities globally. AI systems at modern EAF operations process rendered camera images of the scrap charge bucket contents to classify moisture condition before the bucket is lifted to the charging position: overhead camera views from the scrap bay, side-view infrared camera images of the bucket contents through the bucket side-inspection ports, or a combination of RGB and thermal imaging. The classification categories are: dry/acceptable (no visible water pooling, no ice, scrap surfaces uniformly dry in thermal IR), wet/contaminated (visible water pooling at bucket bottom, ice-covered scrap sections visible, rain-wetted bulk scrap above moisture acceptance threshold), and suspect/marginal (moisture visible on scrap surfaces, condensation in bucket atmosphere, fog or steam visible from scrap mass in cold weather). On wet or suspect classification, the scrap is rejected for charging and returned for drying at the scrap preparation facility.

An adversarial perturbation targeting the wet scrap detection camera AI applies a ±10 DN suppression of visual water indicator pixels in the rendered scrap bucket camera image: shifting the specular reflection signature of pooled water at the bucket bottom (the bright specular highlight characteristic of a water surface under overhead illumination), the crystalline texture signature of ice in scrap cavities (the coherent grain texture of ice as distinct from the rough irregular texture of dry steel scrap), and the condensation signature in the bucket atmosphere (the reduced contrast and haze from water vapour condensation in a cold bucket) to the pixel values characteristic of a dry scrap surface. The EAF scrap bucket AI classifies a wet scrap charge — water pooled at the bucket bottom after overnight rain, ice remaining in the protected cavities of hollow automotive body sections that were stored outdoors in winter — as dry/acceptable. The bucket is loaded onto the overhead charging crane (OSHA 29 CFR 1910.179 crane), the EAF roof is swung open, and the bucket is positioned above the furnace opening. The bucket bottom opens and 80–120 tonnes of scrap, including the pooled water and trapped ice, cascades into the 1,600°C steel bath below. Steam generation occurs in less than 1 millisecond — a water hammer impulse with a peak pressure of hundreds of bar acting on the furnace shell interior — followed by rapid steam expansion: 50 kg TNT equivalent explosion, furnace roof projector, molten steel splash to working area, arc flash from disrupted electrode circuit. The five workers on the EAF operating platform and in the melt shop charging bay — the crane operator, the first operator (EAF pulpit), the second operator (floor), the metallurgist, and the crane signal person — are all within the consequence zone of a steam explosion from a wet scrap charge in a 150-tonne EAF.

3. High-voltage transformer and busbar thermal camera AI (Fluke TiX1000 busbar AI, ABB transformer monitoring AI, Siemens HV thermal management AI — EAF HV transformer and bus duct thermal classification AI)

The EAF furnace transformer — the electrical heart of any electric arc furnace installation — is a custom-engineered high-current transformer rated 40–150 MVA for a 100–200-tonne EAF, transforming the steelworks high-voltage supply (typically 110–220 kV) to the low-voltage, high-current output required to sustain the arc (secondary voltage 400–1,200 V, secondary current 40–150 kA depending on furnace size and tap setting). The transformer is connected to the electrode arms by the short-circuit network: a high-current bus duct (rigid copper or aluminium busbars in an insulated enclosure), flexible water-cooled cables (to accommodate electrode arm movement), delta-connected buswork, mast busbars (the vertical buswork rising inside the electrode support arms), and the electrode contact clamps and electrode arms. All elements of the short-circuit network carry full furnace secondary current (40–150 kA) at every point in the arc cycle — making the I²R heating in every bolted joint, cable termination, and busbar connection a continuous concern. IEC 60076 (Power transformers) specifies maximum temperatures for EAF transformer oil (105°C maximum for ONAN cooling, 95°C continuous operating limit) and windings (125°C maximum for Class B insulation). IEC 61439 (Low-voltage switchgear and controlgear assemblies) specifies maximum busbar surface temperatures of 70°C above ambient temperature for copper busbars in the short-circuit network — a limit equivalent to approximately 90–95°C absolute temperature at typical EAF installation ambient conditions. XLPE-insulated flexible cables in the short-circuit network have a conductor temperature rating of 90°C continuous and 250°C short-circuit emergency. Thermal cameras — including Fluke TiX1000 industrial thermal imagers, ABB transformer monitoring thermal camera systems, and Siemens HV installation thermal management cameras — monitor the transformer oil radiators, busbar joint surfaces, flexible cable sheaths, and delta buswork for developing hot spots during each furnace heat, producing rendered false-colour thermal images that AI systems classify to identify connection overheating and arc fault precursors requiring maintenance intervention before the next tap opportunity.

A thermal hot spot at a bolted busbar joint in the EAF short-circuit network follows a well-documented escalation pathway: an oxidised or loose bolted busbar joint has elevated contact resistance compared to a newly torqued clean joint; the elevated contact resistance (even a 10–50 microohm increase above design at a joint carrying 80 kA) produces additional I²R heating proportional to the square of the current — at 80 kA, a 20 microohm increase in contact resistance produces 128 kW of additional heating at that joint; the additional heating raises the joint temperature above the ambient copper temperature; the elevated temperature accelerates the oxidation of the copper contact surfaces (copper oxide is a semiconductor with orders of magnitude higher resistivity than pure copper), further increasing contact resistance and initiating a thermal runaway feedback: more resistance → more heat → more oxidation → more resistance. At busbar joint temperatures above 200–300°C, the insulation on the adjacent flexible cables begins to degrade; at 400–600°C, the joint may arc internally and the arc fault at full transformer secondary current (40–100 kA available short-circuit current) releases arc flash energy of 200+ cal/cm² at one metre distance — a value far exceeding even the highest PPE rating (NFPA 70E arc flash protection, HRC 4 maximum, 40 cal/cm²). An adversarial perturbation applying a ±8 DN colour shift in the thermal hotspot pixel region of the rendered busbar thermal camera image — shifting the apparent joint temperature from the 85–120°C elevated range (rendered in yellow-orange in the false-colour thermal scale) to the 55–65°C normal range (rendered in blue-green) — causes the EAF busbar thermal AI to classify a developing contact resistance thermal runaway (joint at 90–120°C, above the IEC 61439 limit for copper busbar surface temperature, indicating an oxidised or loose connection) as a within-normal operating condition. The maintenance intervention — re-torquing of the joint bolts to design torque (typically 200–400 Nm for M24–M36 busbar bolts), oxide layer removal with emery cloth and contact compound, or joint replacement — is deferred until the next scheduled maintenance outage. The thermal runaway continues; the joint arc fault initiates at the next high-current tap. EN ISO 13849-1 (Safety of machinery — Safety-related parts of control systems) requires that safety functions in steel processing control systems be implemented with performance level PLe, Category 4 for protection against life-threatening hazards — a requirement that applies to the protective interlock logic driven by EAF transformer and busbar monitoring AI, but which does not specify adversarial robustness requirements for the AI classifying the rendered thermal camera images that feed the protective logic. ArcelorMittal Ghent EAF electrode regulation failures causing sidewall arc attachment and panel burnout — with consequent steel splash and molten metal projection — represent a related consequence pathway from undetected electrical system deviations in the EAF arc circuit.

4. Electrode position deviation camera AI (Danieli QXS electrode regulation AI, Siemens SIMETAL electrode position AI, SMS group Quantum electrode AI — EAF electrode position deviation camera classification AI)

EAF electrode positioning is one of the most demanding control problems in industrial automation: three graphite electrodes — each consisting of a 600–700 mm diameter electrode column assembled from individual sections 1.8–2.5 m long, pin-jointed, with a total hanging mass of 3,000–6,000 kg per electrode assembly — must be maintained at precise vertical positions above the steel bath to sustain arcs of the correct length (arc voltage / arc impedance set-point), with lateral positions centred on the furnace geometry to avoid sidewall arc attachment. The hydraulic electrode regulation system — comprising linear displacement transducers (LDTs) in the electrode cylinder bores measuring electrode arm position, hydraulic servo valves controlling the hydraulic cylinder extension/retraction at regulation cycle times of 50–200 ms, and closed-loop arc current and voltage control by the electrode regulation controller — continuously adjusts electrode position to maintain the arc current and impedance set-points. The Danieli QXS electrode regulation AI, Siemens SIMETAL EAF Optimizer electrode AI, ABB ArcPulsed™ control AI, and Lincoln Electric electrode regulation AI systems use neural network and model-predictive control approaches trained on arc current, voltage, and power waveforms to optimise electrode position set-points in real time. Electrode position deviation — lateral displacement of the electrode tip from the furnace geometric centre axis — occurs from electrode joint failure (a threaded pin joining two electrode sections failing in bending from the mechanical stresses of scrap collapse during the melting phase), electrode arm mechanical failure (hydraulic cylinder damage from a flying refractory fragment during charging), or electrode regulation system fault (a servo valve stuck at a commanded position causing uncontrolled electrode arm travel).

When an electrode tip deviates laterally from its design position — approaching the furnace sidewall water-cooled panels rather than maintaining the centred arc position above the steel bath — the arc attachment migrates from the bath surface to the nearest high-conductivity surface: the WCP interior or the sidewall refractory brick face above the slag line. A sidewall arc attachment produces a localised thermal flux of 5–10 MW/m² on the WCP interior surface, compared to the design basis of 1–2 MW/m² from radiant heat transfer in normal furnace operation. At 5–10 MW/m² localised flux, a steel WCP (thermal conductivity approximately 50 W/m·K, nominal wall thickness 8–12 mm) experiences a surface temperature rise on the arc-impingement face sufficient to initiate melting within 1–5 seconds — far faster than the response time of the manual monitoring cycle. Additionally, an electrode arm displaced 100–200 mm from the furnace centre toward the furnace shell may contact the secondary bus duct structure or the furnace shell steel casing, creating a high-current arc-to-ground fault at full transformer secondary current — with the same 40–100 kA arc flash consequence as the busbar fault pathway described in surface 3 above. AI systems in the EAF electrode monitoring layer process rendered images from electrode position cameras — overhead or side-view optical cameras with clear sightlines to the electrode arms and tips, generating rendered images showing electrode tip position relative to a furnace centre cross-hair reference overlay — to classify electrode alignment: within normal position tolerance (electrode tip within ±50 mm of furnace geometric centre axis), deviated — minor (tip displaced 50–100 mm, increased regulation gain required), deviated — major (tip displaced 100–200 mm, regulation intervention and electrode inspection required), and emergency — sidewall approach (tip within 200–300 mm of furnace sidewall panel, immediate arc power reduction and electrode retraction required).

An adversarial perturbation targeting the electrode position camera AI applies a ±8 DN shift to the pixel region encoding the electrode tip position indicator in the rendered camera image — shifting the apparent electrode tip position from the deviated or emergency range back toward the furnace centre cross-hair reference. The AI classifies a significantly deviated electrode position (electrode tip displaced 150–200 mm from centre toward the furnace sidewall panel, placing arc attachment on WCP interior surface at 5–10 MW/m² localised flux) as within normal operating position tolerance. The electrode regulation system is not commanded to re-centre the electrode; arc power reduction is not initiated; the WCP inspection cycle is not triggered. Sidewall arc attachment continues: the WCP interior wall reaches melting temperature at the arc impingement point within seconds, the panel wall burns through, and cooling water discharges into the furnace interior, initiating the same steam explosion consequence pathway described in surface 1 above — but with the additional complication that the arc remains active at full power during the steam explosion initiation, since the electrode regulation AI has suppressed the emergency retraction command. The ArcelorMittal Ghent EAF electrode regulation failure causing sidewall arc attachment — with consequent panel burnout and steel splash — demonstrates that electrode position deviations of this magnitude occur in real EAF operations. AIST T-13 requires that electrode regulation system failures trigger immediate arc circuit opening (main arc circuit breaker trip) via hardwired relay logic — but does not specify adversarial robustness requirements for AI systems classifying the rendered electrode position camera images that serve as the primary automated indicator of electrode lateral deviation between LDT sensor readings. Free tier — 10 scans/day, no card required.

Integration: EAF AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for steel EAF AI belongs at every rendered-image ingestion boundary in the EAF process monitoring pipeline — before furnace shell WCP thermal camera AI processes rendered thermal false-colour panel images, before wet scrap bucket detection AI processes rendered bucket camera images, before transformer and busbar thermal AI processes rendered short-circuit network thermal images, and before electrode position deviation AI processes rendered electrode position camera images. Threshold 35 for EAF AI contexts reflects the significant life-safety consequences from steam explosion (WCP burnthrough or wet scrap charge: up to 50 kg TNT equivalent, furnace shell rupture, molten steel projection) and arc flash (busbar joint thermal runaway or electrode arc-to-ground fault: 40–100 kA, > 200 cal/cm² incident energy), calibrated against the additional independent detection layers present in EAF operations that are not present in nuclear or hydrogen fuel cell safety architectures: OSHA 29 CFR 1910.179 crane inspection requirements and AIST T-13 physical scrap inspection protocols provide pre-AI manual verification of scrap moisture condition; WCP cooling water flow interlocks (independent hardwired flow and temperature sensors in the WCP cooling water circuit) provide an independent non-AI burnthrough detection layer; and AIST T-13 electrode regulation system fault interlocks provide an independent hardwired arc-trip layer for electrode mechanical failures. These independent layers raise the threshold from the 25–30 used for nuclear I&C and FCEV contexts — where the AI is often the only automated detection layer — to 35, while still requiring pre-scan adversarial robustness verification before any EAF AI classification is accepted as the basis for a safety decision.

import asyncio, base64, hashlib, json
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path

import httpx

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

# EAF AI contexts: threshold 35
# OSHA 29 CFR 1910.179 (Overhead and gantry cranes — scrap bucket operations);
# AIST T-13 (formerly AISE 12.1) EAF Safety and Health Standards;
# EN ISO 13849-1 (Safety-related parts of control systems — PLe, Cat. 4);
# NFPA 484 (Standard for Combustible Metals).
EAF_THRESHOLD = 35


class EAFAIContext(Enum):
    FURNACE_SHELL_THERMAL    = "furnace_shell_thermal"    # WCP hotspot thermal camera AI
    WET_SCRAP_DETECTION      = "wet_scrap_detection"      # Scrap bucket moisture camera AI
    TRANSFORMER_BUSBAR_THERMAL = "transformer_busbar_thermal"  # HV bus thermal camera AI
    ELECTRODE_POSITION       = "electrode_position"       # Electrode deviation camera AI


class AdversarialEAFImageError(Exception):
    """Raised when Glyphward detects adversarial content in an EAF AI
    rendered monitoring image above EAF_THRESHOLD (35).

    Consequence if not raised:
    - FURNACE_SHELL_THERMAL: WCP burnthrough precursor suppressed →
      cooling water contacts 1,600 C bath → steam explosion up to 50 kg
      TNT equivalent → furnace shell rupture → molten steel projection;
      SSAB Oxelösund 2015 WCP burnthrough structural parallel.
    - WET_SCRAP_DETECTION: wet scrap charge classified as dry →
      water in 100-tonne charge contacts bath → steam explosion
      (1–2 L water = 10–20 kg TNT equivalent, < 1 ms) → 40–60% of
      EAF fatalities globally from wet scrap (AIST data).
    - TRANSFORMER_BUSBAR_THERMAL: busbar contact resistance thermal
      runaway suppressed → joint arc fault at 40–100 kA → arc flash
      > 200 cal/cm² → worker fatality / severe burns.
    - ELECTRODE_POSITION: electrode sidewall deviation suppressed →
      sidewall arc attachment at 5–10 MW/m² → WCP burnthrough →
      steam explosion (same consequence as FURNACE_SHELL_THERMAL);
      or electrode arm arc-to-ground fault → 40–100 kA arc flash.
    Fail-safe: halt EAF AI classification; require manual WCP
    thermocouple check (FURNACE_SHELL_THERMAL / ELECTRODE_POSITION),
    physical scrap inspection per AIST T-13 Section 4.2
    (WET_SCRAP_DETECTION), or IR thermography re-scan by qualified
    technician (TRANSFORMER_BUSBAR_THERMAL) before resuming
    AI-driven protective decisions.
    """

    def __init__(
        self,
        scan_id: str,
        score: int,
        context: EAFAIContext,
        plant_id: str,
        furnace_id: str,
        flagged_region: dict | None = None,
    ) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.plant_id = plant_id
        self.furnace_id = furnace_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial EAF image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} furnace={furnace_id} scan_id={scan_id}"
        )


async def scan_eaf_image(
    image_bytes: bytes,
    context: EAFAIContext,
    plant_id: str,
    furnace_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an EAF AI rendered monitoring image for adversarial content.

    Fail-safe contract: AdversarialEAFImageError or httpx error →
    halt EAF AI classification for the affected monitoring context;
    require manual verification per AIST T-13 / OSHA 29 CFR 1910.179
    before resuming AI-driven protective action decisions.

    Args:
        image_bytes: Raw bytes of the rendered EAF monitoring camera image.
        context: EAFAIContext enum value identifying the monitoring surface.
        plant_id: Steelworks plant identifier (e.g. 'nucor-darlington').
        furnace_id: EAF furnace identifier within the plant (e.g. 'eaf-1').
        client: Shared httpx.AsyncClient (connection-pooled, timeout configured).

    Returns:
        Glyphward scan result dict with keys: scan_id, score, flagged_region,
        timestamp_utc. Score below EAF_THRESHOLD: image cleared for EAF AI
        classification. Score at or above EAF_THRESHOLD: raises
        AdversarialEAFImageError — do not pass image to EAF AI classifier.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"eaf:{context.value}:{plant_id}:{furnace_id}",
        "metadata": {
            "plant_id": plant_id,
            "furnace_id": furnace_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"] >= EAF_THRESHOLD:
        raise AdversarialEAFImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            furnace_id=furnace_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_eaf_image at each EAF AI rendered-image ingestion boundary: before furnace shell WCP thermal camera AI (threshold 35), before wet scrap bucket detection camera AI (threshold 35), before transformer and busbar thermal camera AI (threshold 35), and before electrode position deviation camera AI (threshold 35). On AdversarialEAFImageError for WET_SCRAP_DETECTION context: immediately suspend scrap charging operations, initiate physical scrap inspection per AIST T-13 Section 4.2 (visual inspection of bucket contents with direct line-of-sight from a safe distance, supplemented by infrared moisture sensor probe in accessible scrap sections), and do not resume charging AI classification until the physical inspection confirms dry scrap status. On AdversarialEAFImageError for FURNACE_SHELL_THERMAL or ELECTRODE_POSITION: reduce arc power to standby level immediately and initiate manual thermocouple cross-check on the flagged WCP circuit before resuming full-power arc operation. See also: industrial control systems SCADA AI prompt injection (related process control AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions