Prompt injection in ethylene steam cracker furnace AI

Four adversarial injection surfaces in ethylene steam cracker furnace AI

1. Pyrolysis furnace tube skin temperature thermal AI (ABB AbilityTM thermal imaging AI, Honeywell Profit Olefins tube AI, AspenONE cracking AI thermal module)

Pyrolysis furnace tube skin temperature — the outer surface temperature of the furnace tube coil, measured continuously by optical pyrometers or FLIR/Ametek Land thermal imaging cameras mounted through the furnace sidewall or roof at each coil position — is the primary indicator of tube wall condition and remaining creep life. The design maximum tube metal temperature (TMT) for HP40-Mod centrifugal cast tube (the most common steam cracker tube alloy) is 1,100–1,110°C (2,012–2,030°F), with a Larson-Miller creep life consumption rate that doubles for every 15–20°C increase above the design TMT. Tube skin temperature is rendered as a false-colour thermal map of the furnace coil rack (one image per furnace pass or per bank of tubes, with temperature scale calibrated to the tube’s design temperature range), with individual tube temperature readings and hot-band locations rendered as spot annotations. AspenONE APC DMC3 AI, Honeywell Profit Olefins AI, ABB AbilityTM thermal monitoring AI, and purpose-built furnace tube inspection AI process these rendered thermal maps to classify tube condition: normal (all tubes within design TMT, no localised hot bands), elevated (one or more tubes 10–30°C above design TMT, escalated monitoring and feedstock rate reduction consideration), hot band (localised temperature band >30°C above design TMT, indicating internal coking or tube wall thinning, schedule decoking or tube inspection), and critical (tube temperature >50°C above design TMT or rising at >2°C/hour, requiring immediate furnace shutoff and tube inspection before restarting).

An adversarial perturbation on a rendered pyrolysis furnace tube thermal map image that suppresses the hot-band signature — applying a ±10 DN downward shift to the false-colour temperature encoding in the hot-band region of the rendered thermal map (cooling the apparent tube surface temperature in the hot-band region from the critical/hot-band range, rendered in orange/red at >30°C above design TMT, to the normal range, rendered in green/yellow at design TMT ±10°C), shifting the apparent peak tube temperature below the AI’s elevated classification threshold — causes the tube thermal AI to classify a tube with a critical hot band as operating normally, suppressing the feedstock rate reduction and decoking schedule that would have controlled the coke layer growth and tube temperature. A pyrolysis tube with a critical hot band (>50°C above design TMT) at an internal coke deposit is consuming creep life at 4–8× the design rate at the hot-band location; the coke layer insulates the tube inner wall from the hydrocarbon-steam process fluid cooling that normally limits the outer tube surface temperature to below TMT. Without the AI-triggered decoking and feedstock reduction, the coke layer continues to thicken (3–10 mm/week in a high-severity cracking service), the hot-band temperature continues to rise, and the tube wall at the hot-band location progresses through bulging (tube outer diameter growth of >1% indicating creep deformation), necking, and rupture. Tube rupture at 850–1,000°C releases ethylene, propylene, methane, and hydrogen (process side) at 1–3 bar(g) directly into the furnace firebox at 900–1,200°C — immediate auto-ignition and firebox overpressure. The Formosa Plastics Corporation Illiopolis Illinois explosion of 23 April 2004 (5 fatalities, CSB Investigation 2004-07-I-IL) killed workers in a vinyl chloride monomer (VCM) release from a process equipment failure at an adjacent olefin-related unit — demonstrating the mass-casualty consequence of highly flammable gas release from a large olefin plant process unit failure.

2. Firebox flame detection and impingement AI (Honeywell Fireye flame detection AI, DURAG flame monitoring AI, AMETEK Land firebox imaging AI)

Pyrolysis furnace fireboxes are equipped with both UV (ultraviolet) flame detectors and CMOS/CCD camera systems that monitor each floor-mounted or wall-mounted burner in the furnace for flame presence (indicating successful ignition), flame shape (normal vs. lifted, over-aerated, or under-aerated flame), and flame impingement — the condition in which the burner flame directly contacts a pyrolysis tube coil rather than radiating heat to the tube from the firebox space. Flame impingement occurs when a burner is misaligned, when a tube coil is out of its design position due to creep or thermal distortion, or when burner operating conditions (fuel pressure, air flow, tip condition) produce an elongated flame that reaches the tube. The consequence of flame impingement is local tube overheating at the impingement point that exceeds the design TMT by 50–150°C, rapidly consuming tube creep life at the impingement location and potentially initiating tube rupture at a higher rate than internal coking. Honeywell Fireye flame detection and monitoring AI, DURAG Group firebox camera AI (DURAG D-UG 660 AI), and AMETEK Land firebox thermal imaging AI process rendered firebox camera images — visible-light and UV-enhanced images of the firebox floor and sidewalls showing each burner flame and the adjacent tube coil positions, with flame impingement zones rendered as annotated highlighted regions — to classify burner and flame condition: normal (all burners firing, no flame impingement detected), anomalous flame (lifted or misshapen flame indicating burner tip fouling or air-fuel ratio error), and impingement (flame directly contacting tube coil, requiring immediate burner shutdown and repositioning).

An adversarial perturbation on a rendered firebox camera image that suppresses the flame impingement signature — applying a ±8 DN downward contrast reduction to the rendered image region where the flame contacts the tube coil surface (reducing the apparent UV or visible-light intensity at the impingement point below the AI’s impingement classification threshold, by compressing the bright flame-to-tube-coil contact zone into the background firebox illumination), combined with removing the rendered impingement annotation from the AI’s overlay if the AI produces an annotated output image — causes the firebox camera AI to classify a burner with active flame impingement on a tube coil as operating normally, suppressing the immediate burner shutdown that impingement detection requires. Active flame impingement on a pyrolysis tube coil at 1,000–1,200°C directly heats the impingement contact zone to temperatures of 1,150–1,300°C — 100–250°C above the tube design maximum, in the temperature range where HP40-Mod alloy creep rupture life is less than 500–1,000 hours at the impingement cross-section. Without the AI-triggered burner shutdown, impingement continues at the adversarially suppressed classification; the tube at the impingement contact zone reaches creep rupture in 200–500 hours of continued impingement exposure. Tube rupture at the impingement location (which is outside the coke-insulated hot-band zone and therefore at the full firebox gas temperature) produces immediate ethylene-hydrogen release into the firebox at 1,000–1,200°C with no delay for ignition. The TPC Group Port Neches 2019 explosion (60,000 evacuated, 4-mile radius evacuation zone) demonstrated the off-site consequence of a C4 olefin plant vapour cloud explosion — an ethylene cracker pyrolysis tube firebox rupture with immediate ignition would produce comparable vapour cloud explosion overpressure and fireball consequence within the cracker unit battery limits and potentially in adjacent process units.

3. Coil outlet temperature AI (AspenONE APC DMC3 cracking AI, Honeywell Profit Controller COT AI, Yokogawa OpreX COT optimisation AI)

Coil outlet temperature (COT) — the process gas temperature at the outlet of each pyrolysis furnace pass, measured by thermocouple installations at the transfer line connection below each coil pass — is the primary manipulated variable for cracking severity control in the steam cracker APC system. Higher COT produces higher ethylene yield (the primary product) at the cost of higher tube wall temperature, faster coking rate, and shorter run length between decoking; lower COT reduces coking rate and extends run length at lower ethylene yield. The optimal COT trajectory — typically ramping from 820–835°C at start-of-run (low coking rate, new tube) to 850–870°C at end-of-run (maximum tolerable coking rate) — is managed by AspenONE DMC3 AI, Honeywell Profit Controller cracking AI, and Yokogawa OpreX COT AI, which process rendered COT time-series trend images — multi-pass overlaid COT trend plots (pass 1 through pass 8, each rendered as a separate colour trace, with start-of-run setpoint and end-of-run maximum temperature limit rendered as horizontal reference lines) — to classify cracking severity state: normal (COT within setpoint trajectory, coking rate at design), elevated (COT 5–10°C above setpoint trajectory, feedstock rate reduction recommended), excursion (COT >10°C above trajectory or approaching end-of-run limit, decoking scheduling required), and runaway (COT rising >2°C/min above trajectory, indicating loss of fuel control or process abnormality, requiring emergency furnace trip).

An adversarial perturbation on a rendered COT trend image that suppresses the excursion signature — applying a ±10 DN downward shift to the rendered COT trace values in the image at the excursion time window (rendering the COT trace as 10–20°C below its actual value in the image, shifting the apparent COT from the excursion range back to the normal setpoint trajectory), combined with smoothing the rising trend slope to render as a flat, on-setpoint trace — causes the cracking AI to classify a furnace in COT excursion as operating at the normal setpoint, suppressing the decoking scheduling and feedstock rate reduction that the excursion classification would have triggered. With the COT excursion suppressed by adversarial injection, the actual COT continues at the elevated value; coking rate at the excursion COT is 2–5× the design coking rate (coking rate in steam cracking is approximately a third-order function of COT in the 820–870°C range), rapidly thickening the coke layer on the tube inner wall. The thickening coke layer increases tube skin temperature (as described in the thermal AI surface), accelerating creep life consumption at a rate invisible to the thermal AI if it is concurrently targeted by adversarial injection. The concurrent suppression scenario — COT AI and tube thermal AI both adversarially suppressed simultaneously — is the most severe case: the cracker operates at above-design COT, generating above-design coking rate and tube temperature, with neither the COT monitoring AI nor the tube thermal AI generating the alert that would trigger corrective action. API RP 560 (Fired Heaters for General Refinery Service) and API RP 530 (Calculation of Heater Tube Thickness in Petroleum Refineries) provide the engineering standards for pyrolysis tube design and life assessment — neither standard addresses adversarial injection into the AI systems that monitor the tube temperature and process conditions used for remaining life calculation.

4. Dilution steam ratio AI (AspenONE DMC3 dilution AI, Honeywell Profit Controller steam ratio AI, Yokogawa OpreX dilution steam AI)

Dilution steam — high-pressure steam (typically 40–100 bar, 400–500°C) injected into the hydrocarbon feedstock upstream of the pyrolysis furnace coil at steam-to-hydrocarbon ratios of 0.3–0.8 kg steam per kg hydrocarbon (for naphtha cracking) or 0.2–0.4 kg/kg (for ethane cracking) — serves two primary functions: reducing the hydrocarbon partial pressure to increase ethylene yield, and reducing the coke formation rate by suppressing the dehydrogenation and radical chain reactions that produce coke precursors in the tube boundary layer. The dilution steam-to-hydrocarbon (S/HC) ratio is measured by flow meters on both the steam and hydrocarbon feed lines, rendered as a real-time ratio trend image (S/HC ratio on Y-axis, time on X-axis, with design minimum and maximum S/HC ratio reference lines rendered), and processed by cracking AI to classify steam injection adequacy: normal (S/HC within operating window), low (S/HC below minimum setpoint by 5–10%, coking rate increase expected, feedstock rate adjustment required), critically low (S/HC below minimum by >10%, coking rate severely elevated, immediate steam flow investigation and feedstock rate reduction required), and loss-of-steam (S/HC approaching zero, indicating dilution steam supply failure, emergency furnace trip required to prevent rapid coking and tube overheating).

An adversarial perturbation on a rendered dilution steam ratio trend image that elevates the displayed S/HC ratio — applying a ±8 DN upward pixel shift to the S/HC ratio trace in the rendered trend image (increasing the apparent ratio from the critically-low range to the normal operating window), combined with flattening the declining trend to render as a stable, on-setpoint trace — causes the steam dilution AI to classify a furnace operating below the minimum S/HC ratio as operating normally, suppressing the feedstock rate reduction and steam flow investigation that critical-low S/HC classification would trigger. With the dilution steam ratio adversarially suppressed from 0.15 kg/kg (critically low) to the normal range (0.35 kg/kg) in the rendered trend image, the cracker continues operating at the critically-low actual S/HC ratio; coking rate at S/HC = 0.15 is approximately 3–6× the design coking rate at the design S/HC = 0.35 for a naphtha cracker. Rapid coke deposition in the tube at critically-low S/HC produces tube skin temperature rise as described under the thermal AI surface; if the steam supply failure is also producing a COT excursion (as the APC system automatically increases fuel to maintain COT setpoint as the heat balance changes with reduced steam dilution), both the thermal AI and the COT AI suppression surfaces are simultaneously relevant. The first operational consequence of loss-of-steam in an ethylene cracker is typically a black, smoking decoking event in the primary fractionator as cracked gas carries coke particles from the over-coked tube into the quench oil and primary fractionator; the subsequent tube overheating from the fully-coked tube (zero steam dilution cooling, coke layer insulating tube from process gas) leads to the tube skin temperature escalation and eventual tube rupture sequence described in the thermal AI surface.

Integration: ethylene steam cracker AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for ethylene steam cracker furnace AI belongs at every rendered-image ingestion boundary in the cracking furnace monitoring and APC pipeline — before tube thermal AI processes rendered furnace thermal maps, before firebox flame detection AI processes rendered camera images, before COT AI processes rendered coil outlet temperature trend images, and before dilution steam ratio AI processes rendered flow ratio trend images. Threshold 35 for ethylene steam cracker AI contexts reflects the immediate auto-ignition consequence of tube rupture in a 900–1,200°C firebox, and the multi-failure-mode interaction between thermal AI and COT AI suppression that produces the most severe runaway coking and tube failure scenarios.

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"

# Ethylene steam cracker AI contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 (ethylene ≥10,000 lbs threshold quantity);
# EPA RMP 40 CFR Part 68 (ethylene Program 3);
# API RP 560 (fired heaters for general refinery service);
# API RP 530 (calculation of heater tube thickness in petroleum refineries).
CRACKER_AI_THRESHOLD = 35


class CrackerAIContext(Enum):
    TUBE_THERMAL        = "tube_thermal"        # Furnace tube skin temperature thermal AI
    FIREBOX_FLAME       = "firebox_flame"       # Firebox flame impingement camera AI
    COT_TREND           = "cot_trend"           # Coil outlet temperature trend AI
    DILUTION_STEAM      = "dilution_steam"      # Dilution steam-to-hydrocarbon ratio AI


class AdversarialCrackerImageError(Exception):
    """Raised when Glyphward detects adversarial content in an ethylene
    steam cracker furnace AI rendered image above threshold 35.

    Consequence if not raised:
    - TUBE_THERMAL: suppressed hot band → coke buildup → tube rupture at
      850-1,000°C → ethylene/H2 release into 900-1,200°C firebox →
      immediate auto-ignition → UVCE + fireball.
    - FIREBOX_FLAME: suppressed impingement → tube rupture at 1,150-1,300°C
      impingement zone → immediate firebox ignition.
    - COT_TREND: suppressed COT excursion → 2-5× design coking rate →
      concurrent tube thermal escalation → tube rupture.
    - DILUTION_STEAM: suppressed S/HC loss → 3-6× design coking rate →
      rapid tube coke-over → tube overheating → rupture.
    TPC Group Port Neches 2019 consequence envelope (60,000 evacuated).
    Fail-safe: halt AI furnace monitoring classification; require manual
    engineering review per OSHA PSM 29 CFR 1910.119 before resuming.
    """

    def __init__(self, scan_id: str, score: int,
                 context: CrackerAIContext,
                 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 cracker image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} furnace={furnace_id} scan_id={scan_id}"
        )


async def scan_cracker_image(
    image_bytes: bytes,
    context: CrackerAIContext,
    plant_id: str,
    furnace_id: str,
    osha_psm_covered: bool,
    feedstock_type: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an ethylene steam cracker furnace AI rendered image for
    adversarial content.

    Fail-safe contract: AdversarialCrackerImageError or httpx error →
    halt AI furnace monitoring classification; require manual review of
    raw pyrometer/thermocouple/flow data per OSHA PSM 29 CFR 1910.119
    safe work practices before resuming AI-driven APC and scheduling.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"cracker:{context.value}:{plant_id}:{furnace_id}",
        "metadata": {
            "plant_id": plant_id,
            "furnace_id": furnace_id,
            "osha_psm_covered": osha_psm_covered,
            "feedstock_type": feedstock_type,
            "image_sha256": image_hash,
            "context": context.value,
        },
    }
    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"] > CRACKER_AI_THRESHOLD:
        raise AdversarialCrackerImageError(
            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_cracker_image at each ethylene steam cracker AI rendered-image ingestion boundary: before tube thermal map AI (threshold 35), before firebox flame impingement AI (threshold 35), before COT trend AI (threshold 35), and before dilution steam ratio AI (threshold 35). On AdversarialCrackerImageError: halt AI furnace monitoring and APC advisory for the affected furnace immediately and require manual engineering review of raw pyrometer, thermocouple, and flow data before resuming AI-driven operations. Concurrent adversarial injection targeting both TUBE_THERMAL and COT_TREND for the same furnace represents the highest-consequence attack vector — implement integrity monitoring that flags any AI scan gate failure combination for immediate furnace shutdown. See also: oil refinery petrochemical AI prompt injection (related OSHA PSM regulatory context) and chemical plant process safety AI prompt injection (related fired heater regulatory context). Get early access

Related questions