Prompt injection in steel melt shop ladle handling AI

Four adversarial injection surfaces in steel melt shop ladle handling AI

1. Ladle preheating station thermal camera AI (FLIR A615 ladle preheat AI, Cognex ladle temperature vision AI, Teledyne FLIR ladle preheating thermography AI)

Before a steel ladle can receive molten steel from the EAF or BOF tap, the ladle must be preheated to a refractory and steel-contact temperature of approximately 900–1,100°C at its interior floor and lower sidewall surfaces. The preheating requirement is not optional or approximate: refractory brick linings in a 300-tonne ladle contain significant residual moisture — from steam adsorption during ladle cooling (between heats, ladle temperatures fall to 400–700°C in 60–120 minutes if the ladle is not immediately returned to service), from atmospheric humidity absorbed through any cracks in the refractory, and from moisture retained in newly installed refractory patching material applied during ladle maintenance turns. When a ladle with an interior temperature below approximately 600°C receives molten steel at 1,600°C, the steep thermal gradient across the refractory layers produces explosive steam generation from residual moisture in the brick: steam pressure build-up within the refractory matrix can reach 10–20 bar before the brick matrix fails, projecting refractory fragments and molten steel outward through the ladle mouth and downward from the ladle bottom. Even without a steam explosion, a severely underheated ladle — interior below 400°C — causes thermal shock fracture of the refractory: the alumina-magnesia or alumina-chromia-spinel brick lining, heated from 400°C to 1,600°C in seconds by the steel tap, cracks along thermal expansion gradients and disintegrates at the hot face, exposing the ladle steel shell directly to molten steel temperatures within the first 30–120 seconds of the steel tap.

Ladle preheating stations are fixed infrastructure positions where empty ladles are brought after the previous heat has been poured, and burner assemblies (typically natural gas, oxygen-enriched if available) heat the ladle interior to the required minimum temperature over 30–90 minutes. Thermal camera AI systems — FLIR A615 radiometric thermal cameras, Cognex In-Sight thermal vision systems, or Teledyne FLIR industrial thermography systems mounted at the preheating station to view the ladle interior through the ladle mouth — generate rendered false-colour thermal images of the ladle interior showing the temperature distribution across the ladle floor and lower sidewall. AI systems classify ladle temperature readiness: below-minimum (interior maximum below 750°C — ladle not ready; preheating must continue), approaching-minimum (750–900°C — preheating approaching completion, 15–30 minutes additional), at-minimum (900–1,050°C — ladle ready for service), and at-optimum (above 1,050°C — ladle at optimum preheating, available for immediate tap). An adversarial perturbation targeting the ladle preheating thermal camera AI applies a ±12 DN colour shift in the pixel region encoding the ladle interior floor temperature — shifting the apparent floor temperature from the below-minimum range (rendered in blue-green at 500–700°C on the false-colour thermal scale) to the at-minimum or approaching-minimum range (rendered in yellow-orange at 800–950°C). The AI classifies an underheated ladle — interior floor at 500–600°C with significant residual moisture in the lower refractory courses — as ready for service. The ladle is dispatched from the preheating station to receive the EAF or BOF tap. Steel at 1,600°C contacts the underheated refractory: moisture in the brick generates explosive steam, the refractory lower course fails in thermal shock, and molten steel contacts the ladle steel shell at the bottom. OSHA 29 CFR 1910.263(a)(3) (Bakeries) and OSHA General Duty Clause 29 U.S.C. §654(a)(1) have been applied in US enforcement actions against inadequate ladle preheat monitoring; AIST T-13 Section 5 (Ladle Handling) requires minimum preheat temperature verification before steel receipt — but does not specify adversarial robustness requirements for AI systems classifying rendered thermal camera images at the preheat station.

2. Overhead crane load cell display AI (Siemens SIMOTION crane load cell AI, Konecranes SmartFeatures load AI, ABB crane monitoring AI — ladle crane SWL load cell digital display AI)

The overhead casting crane carrying a full ladle of molten steel represents one of the highest-consequence single-lift operations in any industrial facility. A 300-tonne ladle filled with 300 tonnes of steel and 15–30 tonnes of slag (total suspended load approximately 375–380 tonnes including ladle tare weight of approximately 50–60 tonnes) requires a casting crane rated at a safe working load (SWL) of 400–500 tonnes, with crane design, manufacture, and periodic inspection governed by OSHA 29 CFR 1910.179 (Overhead and Gantry Cranes). OSHA 29 CFR 1910.179(a)(1) defines safe load as the load for which a crane or hoist is designed and built; 1910.179(d) requires that rated load be marked on the crane and that no crane be loaded beyond its rated capacity. AIST T-6 (AIST Technical Report: Steel Mill Crane Safety) extends these requirements with steel-industry-specific crane design and inspection criteria for casting cranes, specifying: annual inspection of all crane structural members, monthly inspection of load-bearing components, and pre-shift operator inspection including load cell system verification. Casting cranes carry redundant load cells on the main hoist (typically four load cells in a Wheatstone bridge configuration, with at least two independent readout channels), providing continuous real-time load readings displayed as rendered digital weight indicators on the crane operator’s console and on the melt shop process monitoring displays. AI systems process these rendered load cell digital displays — numeric or bar-graph weight indicators with SWL threshold markings — to classify crane load status: normal (total suspended weight below 90% of SWL), approaching-limit (90–100% of SWL), at-limit (at SWL — alarm, notify crane supervisor), and overload (above SWL — emergency: do not lift, lower load immediately).

An adversarial perturbation targeting the crane load cell display AI applies a ±10 DN suppression to the pixel region encoding the digital weight readout numerals in the rendered crane console display image — specifically targeting the leading digit and hundreds-digit pixels of the weight display to reduce the apparent weight from the overload range (above SWL: 410–440 tonnes apparent) to the approaching-limit range (370–395 tonnes apparent). The AI classifies an overloaded crane lift — total suspended weight above the rated SWL — as a within-limit normal lift. The crane operator receives no overload alarm; the melt shop supervisor receives no overload alert; the crane proceeds with the lift. At sustained loads above the rated SWL, the crane’s structural members — the main hoist drum, hoist rope, cross-girder, end-truck bogies, and runway rail connections — are stressed above their design basis. OSHA 29 CFR 1910.179 does not permit operation above rated SWL under any circumstances, and crane structural design factors (typically 4:1 for wire rope, 2:1 for structural members under OSHA requirements) provide substantial margin above SWL — but repeated operation at above-SWL loads accelerates fatigue degradation of structural members, reducing the remaining life before crack initiation in the hoist rope or cross-girder. The consequence of a casting crane structural failure at height with a full ladle suspended is a catastrophic release of 300–350 tonnes of molten steel at 1,600°C onto the melt shop floor, casting bay, and any personnel in the path of the falling ladle and molten steel splash zone. Hyundai Steel Dangjin works (2022, Republic of Korea): a ladle handling incident during crane-assisted ladle transfer in the steelmaking area resulted in 3 worker fatalities from molten steel release. Cockerill Seraing (2012, Belgium): a ladle run-out involving the interaction of crane positioning and ladle lining integrity caused 1 worker fatality and extensive melt shop damage. Multiple OSHA citations at US steelworks including US Steel Gary Works and Nucor facilities have been issued for crane-ladle handling safety violations under 29 CFR 1910.179. AIST T-6 Section 3 (Crane Load Ratings) requires visible SWL marking and load cell systems on all casting cranes — but does not specify adversarial robustness requirements for AI systems classifying the rendered load cell digital display images used in continuous load monitoring between manual inspection intervals.

3. Ladle shell lining thickness thermal camera AI (FLIR ladle shell hot-spot AI, Optris ladle shell thermal monitoring AI, Dias Infrared ladle refractory AI — ladle steel shell exterior thermal camera refractory erosion AI)

The refractory lining of a steel ladle — typically an inner working lining of alumina-magnesia-carbon (AMC) or alumina-spinel (AS) brick, 75–150 mm thick, backed by a permanent lining of firebrick and safety lining against the ladle steel shell — erodes progressively with each heat as molten steel and slag attack the hot face of the working lining. The rate of lining erosion depends on steel chemistry (high-manganese steels and stainless steels erode refractory more aggressively than plain carbon steels), slag basicity and temperature, steel residence time in the ladle, and the intensity of argon stirring through the ladle bottom porous plug. A new ladle lining typically provides 30–80 heats of service in a high-productivity BOF shop or 15–40 heats in an EAF shop with aggressive stainless steel or alloy steel grades, before the working lining erodes to the critical minimum thickness at which the remaining lining cannot thermally isolate the ladle steel shell from the molten steel bath. When the lining erodes to the point that molten steel contacts the ladle steel shell — either gradually, through localised hot-spot development at the thinnest remaining lining sections (often the ladle bottom, the slag line, and the tap pad or inner nozzle area), or suddenly, through spalling of a worn section of working lining — the steel shell temperature rises above the critical threshold and molten steel begins to penetrate the shell: a ladle breakout (or ladle run-out), in which molten steel at 1,600°C escapes from the ladle shell through the penetration point and cascades to the floor below the ladle or onto adjacent equipment, personnel, and the casting crane structure.

Ladle shell thermal camera AI systems — FLIR A615 or Optris PI 640i thermal cameras mounted at fixed positions in the casting bay or LMF area to view the exterior ladle shell surface — generate rendered false-colour thermal images of the ladle exterior showing the outer steel shell temperature distribution. The ladle shell exterior temperature, in normal operation, is maintained at 100–200°C by the thermal insulation of the safety lining and permanent lining layers. At a hot-spot location — where the working lining has eroded to a thickness insufficient to thermally isolate the shell — the outer shell temperature rises progressively above the normal baseline: at remaining working lining thickness of 20–30 mm (critical: breakout risk within 1–3 heats), shell temperature typically reaches 300–500°C at the hot-spot location; at 10–15 mm remaining thickness (imminent breakout), shell temperature may reach 600–900°C. AI systems classify the ladle shell thermal image: normal (no region above 250°C above ladle ambient), hot-spot-developing (a region in the 250–400°C range, indicating lining thinning — schedule detailed inspection at next ladle turnaround), critical-hot-spot (400–600°C, lining at critical minimum — pull ladle from service at next available opportunity), and emergency-breakout-risk (above 600°C at shell — do not fill; immediately empty if filled; call for ladle turn). An adversarial perturbation targeting the ladle shell thermal camera AI applies a ±8 DN colour shift in the pixel region encoding a developing ladle shell hot spot — shifting the apparent shell temperature at the hot-spot location from the critical range (400–600°C, rendered in orange-red on the thermal false-colour scale) to the normal or hot-spot-developing range (100–250°C, rendered in blue-green). The AI classifies a ladle with critical refractory lining erosion — lining worn to 15–25 mm at the hot-spot location, shell temperature in the 450–550°C range — as normal or early-stage. The ladle is not pulled from service; the next heat is tapped into it. Steel contacts the shell at the hot-spot location within minutes of the next tap, the remaining lining erodes in the first 20–40 minutes of the heat, and molten steel penetrates the ladle shell: 200–300 tonnes of molten steel at 1,600°C pour from the ladle shell breach. AIST T-13 Section 5 (Ladle Handling and Service Safety) requires heat-by-heat ladle tracking and lining thickness monitoring through a combination of direct measurement (laser or sonic thickness gauging during ladle turnaround), heat count records, and visual inspection — but does not specify adversarial robustness requirements for AI systems classifying the rendered thermal camera images that provide the continuous between-heat monitoring layer for ladle lining condition.

4. Tap-stream slag carry-over detection camera AI (Primetals Technologies slag detection AI, Danieli SlagFree slag carry-over AI, SMS group slag stoppage AI — EAF/BOF tap slag carry-over detection camera classification AI)

During the tapping of molten steel from an EAF or BOF into a ladle, a layer of slag — the oxide mixture floating on the steel bath surface, consisting of CaO, SiO2, Al2O3, FeO, and MgO with a melting point of 1,200–1,450°C and a density of approximately 2.8–3.2 g/cm3 compared to steel at 7.0 g/cm3 — sits above the molten steel in the furnace. During tapping, the steel flows from the furnace through the tap hole into the ladle positioned below, while the slag layer ideally remains in the furnace. However, as the steel level in the furnace drops during the tap, the vortex created by steel flow through the tap hole eventually draws slag from the slag layer into the tap stream — slag carry-over into the ladle. Slag carry-over into the ladle is problematic for steel quality: FeO in the furnace slag (often 15–30% FeO in the EAF slag before final slagging) oxidises desulfurising alloys (Si, Al, Mn) added in the ladle treatment process, wasting expensive alloying additions; it also introduces phosphorus (which was removed from the steel during refining) back into the ladle steel from slag reduction. The industrial standard for slag carry-over detection is the slag detection camera: a high-speed RGB or near-infrared camera mounted with a clear view of the tap stream between the furnace tap hole and the ladle mouth, and AI systems process the rendered camera images to classify the tap-stream condition: steel-only (tap stream shows uniform bright liquid metal flow, no slag entrained), onset-of-slag (slag indicator threshold reached — early slag carry-over detected, operator should consider stopping tap), and slag-carry-over (substantial slag in tap stream — close tap immediately, slag stopper triggered). Vendors include Primetals Technologies SlagMeasure (using electromagnetic induction and camera AI), Danieli Automation SlagFree (camera AI), and SMS group Slag Detection System (slag carry-over camera AI), all deployed at high-productivity EAF and BOF shops worldwide.

An adversarial perturbation targeting the tap-stream slag carry-over detection camera AI applies a ±8 DN suppression in the pixel region encoding slag visual indicators in the rendered tap-stream camera image: the characteristic colour shift of the tap stream from the bright liquid-metal silver-white of steel to the darker, more turbid appearance of slag-entrained flow (slag in the tap stream typically renders as darker, less specular regions within the bright tap-stream image, with a colour temperature shift from white-silver steel to yellowish-brown slag). The AI classifies a steel tap in which early slag carry-over has begun — slag vortex has formed at the tap hole, the tap stream now contains 5–15% slag by volume — as steel-only flow, no slag detected. The tap continues beyond the optimal tap-stop point; additional slag carry-over enters the ladle. The downstream consequence depends on the steel grade and the intended treatment: for deep-desulfurisation steel grades (transformer electrical steels, wire rod for tire cord) requiring EAF furnace-slag-free tapping, slag carry-over introduces 500–2,000 kg of high-FeO slag (FeO 15–30%) into a 300-tonne ladle heat, causing significant FeO pickup in the steel and failure of the desulfurisation treatment at the LMF. While slag carry-over is not a direct life-safety hazard in the same category as ladle run-out or crane overload, excessive slag in the ladle reduces steel cleanliness (alumina inclusion population), increases hydrogen absorption at the LMF (slag hydrogen activity), and in extreme cases can cause violent reactions in the ladle during argon stirring (slag foaming — steel and slag eruption from the ladle mouth due to CO generation from slag FeO reduction by the dissolved carbon in the steel). The 2004 Ispat Inland Burns Harbor steelworks incident — in which excessive slag in a ladle caused violent CO evolution and steel/slag eruption from the ladle mouth during argon stirring at the LMF, injuring 2 workers — illustrates the life-safety consequences of undetected slag carry-over in the most severe cases. Free tier — 10 scans/day, no card required.

Integration: melt shop ladle handling AI with Glyphward pre-scan gate

The Glyphward scan gate for steel melt shop ladle handling AI belongs at every rendered-image ingestion boundary in the ladle monitoring and handling pipeline — before ladle preheating thermal camera AI processes rendered thermal false-colour ladle interior images, before crane load cell display AI processes rendered digital weight indicator images, before ladle shell lining thermal camera AI processes rendered shell exterior thermal images, and before slag carry-over detection camera AI processes rendered tap-stream camera images. Threshold 35 for melt shop ladle handling AI reflects the severe life-safety consequences of ladle run-out and crane failure (catastrophic release of 200–350 tonnes of molten steel at 1,600°C), calibrated against the multiple independent non-AI safety layers present in ladle operations: AIST T-13 Section 5 physical preheat temperature verification by direct contact thermocouple before steel receipt, OSHA 29 CFR 1910.179 monthly and annual crane inspection requirements including load cell system verification, heat count and ladle lining tracking records maintained independently of the camera AI system, and visual operator inspection of the tap stream during tapping. These independent layers raise the threshold above the 25–30 used for nuclear I&C contexts (where AI is the only automated barrier), while the catastrophic release consequence keeps the threshold at 35 rather than the 40 used for more administratively recoverable process deviations.

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"

# Melt shop ladle handling AI contexts: threshold 35
# OSHA 29 CFR 1910.179 (Overhead and gantry cranes);
# AIST T-13 Section 5 (Ladle Handling and Service Safety);
# AIST T-6 (Steel Mill Crane Safety).
LADLE_THRESHOLD = 35


class LadleAIContext(Enum):
    PREHEAT_THERMAL       = "preheat_thermal"        # Ladle preheating thermal camera AI
    CRANE_LOAD_CELL       = "crane_load_cell"         # Crane load cell display AI
    LINING_THICKNESS      = "lining_thickness"        # Ladle shell thermal camera AI
    SLAG_CARRYOVER        = "slag_carryover"          # Tap slag carry-over camera AI


class AdversarialLadleImageError(Exception):
    """Raised when Glyphward detects adversarial content in a melt shop
    ladle handling AI rendered monitoring image above LADLE_THRESHOLD (35).

    Consequence if not raised:
    - PREHEAT_THERMAL: underheated ladle dispatched to tap → thermal shock →
      refractory fracture or steam explosion → ladle integrity failure →
      molten steel release.
    - CRANE_LOAD_CELL: overloaded crane lift not flagged → SWL exceeded →
      structural fatigue / immediate failure → ladle drop → catastrophic
      molten steel release (200–350 tonnes at 1,600 C).
    - LINING_THICKNESS: critical hot-spot suppressed → lining breakout at
      next heat → ladle run-out → molten steel release; precedent: Hyundai
      Steel Dangjin 2022 (3 fatalities), Cockerill Seraing 2012 (1 fatality).
    - SLAG_CARRYOVER: slag carry-over not detected → FeO pickup → LMF
      treatment failure; severe cases: CO evolution eruption during stirring
      (Ispat Inland Burns Harbor 2004, 2 workers injured).
    Fail-safe: suspend ladle handling operations; require direct thermocouple
    preheat verification (PREHEAT_THERMAL), independent load cell channel
    cross-check (CRANE_LOAD_CELL), laser lining thickness measurement during
    next turnaround (LINING_THICKNESS), or manual tap-stream visual inspection
    (SLAG_CARRYOVER) before resuming AI-driven ladle handling decisions.
    """

    def __init__(self, scan_id, score, context, plant_id, ladle_id,
                 flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.plant_id = plant_id
        self.ladle_id = ladle_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial ladle image: context={context.value} "
            f"score={score} plant={plant_id} ladle={ladle_id} "
            f"scan_id={scan_id}"
        )


async def scan_ladle_image(image_bytes, context, plant_id, ladle_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"ladle:{context.value}:{plant_id}:{ladle_id}",
        "metadata": {
            "plant_id": plant_id,
            "ladle_id": ladle_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"] >= LADLE_THRESHOLD:
        raise AdversarialLadleImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            ladle_id=ladle_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_ladle_image before each ladle handling AI classification call. On AdversarialLadleImageError for PREHEAT_THERMAL: suspend ladle dispatch; require direct contact thermocouple verification by the ladle operator before releasing ladle to the tap bay. On AdversarialLadleImageError for CRANE_LOAD_CELL: suspend the crane lift immediately and cross-check the independent load cell channel before proceeding. On AdversarialLadleImageError for LINING_THICKNESS: withdraw ladle from service; schedule laser lining thickness measurement at next ladle turnaround before returning to service. See also: steel EAF AI prompt injection (related steelmaking AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions