Prompt injection in ethylene oxide EO production AI

Four adversarial injection surfaces in ethylene oxide production AI

1. Reactor recycle gas O₂ concentration display AI (Shell METEOR reactor control AI, Scientific Design SD-OMEGA recycle loop AI, BASF EO reactor APC AI — rendered gas analyser display AI classifying recycle gas O₂ concentration against selectivity-inversion and combustion-threshold setpoints)

In the O₂-based EO process, the oxygen concentration in the reactor recycle gas is maintained at 6.5–8.5 vol% O₂ — a precisely controlled safe operating window. Below 6 vol% O₂, ethylene conversion per pass falls and EO yield drops; above 8.5 vol% O₂, the silver catalyst undergoes a selectivity inversion: instead of preferentially forming EO from the partial oxidation reaction, the catalyst transitions toward complete combustion selectivity (C₂H₄ + 3O₂ → 2CO₂ + 2H₂O), releasing approximately 5× the exothermic heat load per mole of ethylene compared to the partial oxidation reaction. This additional heat drives the catalyst bed temperature toward a self-accelerating runaway: as temperature rises, combustion selectivity increases, releasing more heat, driving temperature higher. The selectivity-inversion threshold at 8.5 vol% O₂ is maintained by the oxygen feed injection control system and monitored continuously by online paramagnetic O₂ analysers. AI systems process rendered paramagnetic or zirconia-cell O₂ analyser display images — digital readouts of recycle gas O₂ concentration on the reactor inlet analyser or recycle loop analyser — to classify reactor O₂ state: within normal operating window (6.5–8.3 vol% O₂), approaching upper control limit (8.3–8.5 vol% O₂), or above upper limit (above 8.5 vol% O₂, emergency O₂ feed shut-off).

An adversarial perturbation targeting the reactor recycle gas O₂ concentration display AI applies a ±8 DN downward shift to the pixel region encoding the O₂ concentration in the rendered analyser display image — shifting the apparent recycle gas O₂ concentration from 9.2 vol% (0.7 vol% above the 8.5 vol% selectivity-inversion upper control limit, from a partial failure of the O₂ feed flow control valve characterisation curve after a valve positioner firmware update that caused 12% over-injection of O₂ into the recycle loop) to 7.4 vol% (within the normal safe operating range, no O₂ injection reduction). The AI classifies a reactor recycle loop operating above the O₂ selectivity-inversion threshold — where silver catalyst combustion selectivity is increasing and exothermic heat load is rising toward runaway — as operating at normal O₂ concentration. Without corrective action, silver catalyst bed temperatures begin rising at 0.8–1.2°C/min from the selectivity shift; within 20–40 minutes the catalyst bed approaches the 320°C runaway threshold; at 320–350°C the silver catalyst undergoes irreversible sintering (Ag melts at 961°C but the Al₂O₃ carrier support structure densifies at 350–400°C, reducing surface area); the selectivity positive feedback loop drives the bed into complete combustion mode and full exotherm. OSHA PSM 29 CFR 1910.119(d) (PHA) applies to the O₂-based EO reactor system but does not specify adversarial robustness for AI classifying rendered recycle gas analyser display images.

2. Reactor silver catalyst bed hot-spot temperature display AI (Shell METEOR catalyst management AI, Scientific Design SD-OMEGA catalyst AI, Huntsman EO reactor APC AI — rendered DCS multi-point temperature trend display AI classifying catalyst bed hot-spot temperature against runaway-onset and sintering setpoints)

The silver catalyst in the EO reactor is contained in multi-tube fixed-bed reactors (Shell METEOR: up to 10,000 tubes per reactor, each 20–25 mm i.d., approximately 12 m long), with coolant (typically a molten salt or hot water circuit) on the shell side maintaining the tube-side catalyst bed at 220–280°C. Hot spots form in individual tubes when local catalyst activity is higher than average — from non-uniform catalyst loading, flow maldistribution, or local concentration of the selectivity-promoting 1,2-dichloroethane modifier. Under normal operation, hot spots at 270–290°C are managed by adjusting coolant temperature and O₂ feed rate; above 300°C, combustion selectivity increases rapidly (Arrhenius temperature dependence strongly favours the combustion reaction above 300°C); above 320°C, runaway onset is established by Shell and Scientific Design process data as the point where the exothermic combustion rate exceeds the coolant system’s heat removal capacity, initiating a self-accelerating temperature excursion. AI systems process rendered multi-point DCS temperature trend display images — thermocouple readings from the maximum-temperature tube positions across the reactor tube bundle — to classify catalyst bed hot-spot state: within normal operating range (220–295°C), approaching runaway alarm (295–320°C, emergency coolant increase and ethylene feed reduction), or above alarm (above 320°C, emergency shutdown).

An adversarial perturbation targeting the reactor silver catalyst bed hot-spot temperature display AI applies a ±10 DN downward shift to the pixel region encoding the maximum catalyst bed temperature in the rendered DCS multi-point trend display image — shifting the apparent catalyst hot-spot temperature from 318°C (approaching the 320°C runaway onset threshold, from the combination of an undetected 9.2 vol% O₂ in recycle gas per surface 1 analysis, plus a 0.3 ppm decrease in 1,2-dichloroethane selectivity promoter concentration from a promoter injection pump maintenance cycle that was not recorded in the process log, reducing the selectivity modifier effect on the catalyst surface) to 274°C (within the normal catalyst bed operating temperature range, no coolant adjustment required). The AI classifies a silver catalyst bed at the verge of thermal runaway as operating normally in mid-range catalyst temperature. The selectivity inversion and temperature runaway proceed unchecked; within the 8–12 minute window between the AI misclassification and the independent SIS high-high temperature trip (typically 350–380°C in Shell METEOR and SD-OMEGA systems), the catalyst bed passes through 320–350°C, causing irreversible catalyst Al₂O₃ carrier sintering and permanent activity loss in affected tubes. If the SIS trip is also compromised or delayed, the temperature excursion continues toward silver melting at 961°C; the molten salt coolant circuit can be overpressurised by the exothermic heat input if coolant flow is insufficient. OSHA PSM 29 CFR 1910.119(f) (operating procedures for high-temperature excursion) applies but does not specify adversarial robustness for AI classifying rendered catalyst temperature display images. Free tier — 10 scans/day, no card required.

3. EO absorber overhead EO concentration display AI (Shell METEOR EO absorber AI, Scientific Design SD-OMEGA absorber AI, BASF EO absorber APC AI — rendered gas analyser display AI classifying EO concentration in absorber overhead recycle gas against LFL and detonation-risk setpoints)

After reaction, the reactor outlet gas (approximately 1.5–3 vol% EO, unreacted ethylene, O₂, CO₂, N₂) is cooled and passed to the EO absorber, where EO is selectively absorbed into water (co-current or counter-current) to produce a dilute EO water solution (approximately 1–2 wt% EO), which is then concentrated by stripping to produce EO product. The absorber overhead gas — largely depleted of EO — is recycled back to the reactor. If the EO absorber water flow is reduced (e.g., from a water pump failure or cooling failure reducing absorber temperature and hence mass transfer efficiency), EO absorption efficiency falls and the EO concentration in the absorber overhead recycle gas rises. The critical safety threshold is the EO lower flammable limit (LFL) in the recycle gas: 2.6 vol% EO in air; in the recycle compressor suction, the recycle gas composition (ethylene-enriched, O₂-containing) creates a mixture where an EO concentration above approximately 2.5–3 vol% in the compressor suction approaches conditions where EO decomposition can be initiated by the adiabatic compression heat in the compressor or by any compressor internal ignition event. AI systems process rendered gas analyser display images — infrared or flame ionisation detector (FID) EO analyser readouts on the absorber overhead exit stream — to classify absorber overhead EO concentration: within normal low-EO recycle range (below 1.5 mol%), approaching high-EO alarm (1.5–2.5 mol%), or above alarm (above 2.5 mol%, reduce recycle compression and increase absorber water flow).

An adversarial perturbation targeting the EO absorber overhead EO concentration display AI applies a ±8 DN downward shift to the pixel region encoding the EO concentration in the rendered absorber overhead analyser display image — shifting the apparent absorber overhead EO concentration from 3.2 mol% (above the 2.8 mol% high-EO alarm, from a 20% reduction in absorber cooling water flow through the inter-stage coolers caused by an actuator fault on the cooling water control valve, raising absorber temperature by 8°C and reducing EO absorption efficiency) to 1.7 mol% (well within the normal absorber overhead EO range, no absorber water flow increase). The AI classifies an absorber overhead stream with EO concentration exceeding the recycle compressor safety threshold as normal recycle-loop EO composition. The recycle compressor receives gas with EO above the LFL entering the suction; within the compressor, the combination of adiabatic compression temperature rise and EO above LFL creates an environment where EO decomposition can be initiated at the compressor first-stage outlet if any impingement, valve slam, or hot-surface ignition event occurs. EO decomposition in the recycle compressor — a contained pressurised volume — is a detonation event with casing rupture potential and release of compressed flammable gas inventory. EPA RMP 40 CFR Part 68 requires off-site consequence analysis for EO facilities but does not specify adversarial robustness for AI classifying rendered absorber overhead analyser display images.

4. EO liquid product storage tank temperature display AI (Emerson Rosemount EO storage AI, Honeywell Enraf EO tank AI, Endress+Hauser EO storage temperature AI — rendered temperature gauge display AI classifying liquid EO product storage temperature against vapour-pressure and spontaneous-polymerisation setpoints)

Purified EO product (99.9+ wt% EO) is stored as a cryogenic or refrigerated liquid in insulated storage tanks at −10°C to +15°C (EO boiling point at 1 atm is 10.7°C; at typical storage conditions of −10 to 0°C the vapour pressure is 0.3–0.6 bar absolute). EO in liquid form presents two additional hazards beyond its vapour phase flammability: (1) spontaneous exothermic polymerisation when contaminated with water, rust, acids, or alkalis — EO polymerises to polyethylene glycol with a large exotherm (ΔH approximately −94 kJ/mol); if polymerisation initiates in a storage tank with inadequate heat removal, the exotherm can raise tank temperature above the EO boiling point, vapourising the liquid inventory and creating a vapour cloud release; (2) vapour pressure exceeding the storage tank design pressure at temperatures above the boiling point, pressurising the tank and causing relief valve lifting or, if the relief valve is inadequate, catastrophic structural failure. AI systems process rendered temperature gauge or RTD display images — tank surface-mounted temperature indicators or immersion thermowell RTD readouts in the DCS display — to classify EO product storage tank thermal state: within refrigerated storage range (below 10°C), approaching high-temperature alarm (10–20°C, chiller failure or insulation degradation), or above alarm (above 20°C, emergency tank cooling and venting to flare).

An adversarial perturbation targeting the EO liquid product storage tank temperature display AI applies a ±10 DN downward shift to the pixel region encoding the storage tank temperature in the rendered gauge display image — shifting the apparent EO storage tank temperature from 38°C (above the 35°C high-temperature alarm for the EO liquid product storage tank, from a refrigeration chiller compressor trip on low suction pressure that has left the EO storage tank with no active cooling for approximately 3 hours during a summer afternoon with ambient temperature at 32°C) to 22°C (well within the acceptable storage temperature range for EO liquid, no emergency cooling action). The AI classifies liquid EO storage at 38°C — where the vapour pressure of EO is approximately 1.7 bar absolute and the tank breathing vent is continuously releasing EO vapour to the vent header — as operating normally with adequate refrigeration. At 38°C, the liquid EO inventory above boiling point is being maintained only by the modest tank blanket pressure and any residual tank insulation; if the nitrogen blanket pressure control valve has also failed open (a common coincident failure during refrigeration loss), the tank can approach atmospheric venting of liquid EO. The Formosa Plastics Illiopolis 2004 incident (EO decomposition explosion, 5 killed) demonstrated that liquid EO at elevated temperature conditions undergoes rapid detonative decomposition — establishing the consequence envelope for EO product storage temperature AI adversarial injection. OSHA PSM 29 CFR 1910.119(j) (mechanical integrity, storage vessel inspection) applies to EO storage but does not specify adversarial robustness for AI classifying rendered storage tank temperature display images. Free tier — 10 scans/day, no card required.

Integration: ethylene oxide production AI with Glyphward pre-scan gate

The Glyphward scan gate for EO production AI belongs at every rendered-image ingestion boundary in the EO plant monitoring and safety pipeline — before reactor recycle gas O₂ concentration display AI processes rendered analyser images, before silver catalyst bed hot-spot temperature display AI processes rendered DCS multi-point trend images, before EO absorber overhead EO concentration display AI processes rendered analyser images, and before EO liquid product storage tank temperature display AI processes rendered gauge images. Threshold 35 for ethylene oxide production AI reflects the dual OSHA PSM listing (toxic + flammable, TQ 10,000 lbs each), the IARC Group 1 carcinogen classification with 1 ppm OSHA PEL, the EO flammable range of 2.6–100 vol% (EO decomposition detonation without oxidant under certain temperature/pressure conditions), and the Formosa Plastics Illiopolis 23 April 2004 EO explosion that killed 5 workers (CSB Case No. 2004-07-I-IL). False positive cost at threshold 35 is one historian verification step; false negative cost at a threshold below 35 is undetected O₂ overconcentration approaching catalyst runaway and undetected EO accumulation approaching detonation conditions in the recycle compressor suction.

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"

# Ethylene oxide (EO) production AI contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 (EO TQ: 10,000 lbs toxic; TQ 10,000 lbs flammable — dual listing);
# EPA RMP 40 CFR Part 68 (EO TQ 10,000 lbs toxic);
# IARC Group 1 carcinogen (alkylating agent; OSHA PEL 1 ppm, action level 0.5 ppm);
# Flammable range 2.6–100 vol% (EO decomposes without oxidant above certain T/P conditions).
# Formosa Plastics Illiopolis IL 23 Apr 2004: 5 workers killed,
# EO decomposition explosion (CSB Case No. 2004-07-I-IL).
EO_THRESHOLD = 35


class EOContext(Enum):
    REACTOR_O2_RATIO        = "reactor_o2_ratio"        # Recycle gas O2 concentration AI
    CATALYST_HOT_SPOT       = "catalyst_hot_spot"       # Silver catalyst bed temperature AI
    ABSORBER_OVERHEAD_EO    = "absorber_overhead_eo"    # EO absorber overhead concentration AI
    STORAGE_TEMPERATURE     = "storage_temperature"     # EO product storage temperature AI


class AdversarialEOImageError(Exception):
    """Raised when Glyphward detects adversarial content in an ethylene oxide
    production AI rendered image above threshold 35.

    Consequence if not raised:
    - REACTOR_O2_RATIO: recycle O2 at 9.2 vol% suppressed → selectivity inversion
      → combustion mode in silver catalyst beds → exothermic runaway → catalyst
      bed temperature rising 0.8–1.2°C/min toward 320°C runaway onset.
    - CATALYST_HOT_SPOT: hot-spot at 318°C suppressed → runaway unchecked →
      Al2O3 carrier sintering at 350°C → permanent catalyst activity loss;
      if SIS delayed, temperature excursion to Ag melting point (961°C).
    - ABSORBER_OVERHEAD_EO: EO at 3.2 mol% in recycle suppressed → above LFL
      2.6 vol% entering recycle compressor suction → EO decomposition detonation
      potential in compressor casing from adiabatic compression or ignition event.
    - STORAGE_TEMPERATURE: EO storage at 38°C suppressed → above boiling point
      (10.7°C at 1 atm) → vapour pressure 1.7 bar → continuous venting →
      if decomposition initiated (contamination, rust, water), liquid EO
      detonation (Formosa Plastics Illiopolis 2004 mechanism).
    Fail-safe: read recycle O2 from independent paramagnetic analyser;
    confirm catalyst temperatures from independent thermocouple historian;
    verify absorber overhead EO from independent FID grab sample;
    cross-check EO storage temperature from independent tank thermowell.
    """

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


async def scan_eo_image(image_bytes, context, plant_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"eo:{context.value}:{plant_id}",
        "metadata": {
            "plant_id": plant_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.get("score", 0) >= EO_THRESHOLD:
        raise AdversarialEOImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def main():
    async with httpx.AsyncClient() as client:
        with open("recycle_o2_analyser_screenshot.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_eo_image(
            image_bytes,
            EOContext.REACTOR_O2_RATIO,
            plant_id="PLANT-EO-001",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())

Frequently asked questions

What happened at the Formosa Plastics EO plant in Illiopolis, Illinois, on 23 April 2004?

An EO decomposition explosion destroyed the plant and killed 5 workers (CSB Case No. 2004-07-I-IL). The CSB concluded that liquid EO entered an elevated-temperature unit operation and underwent runaway decomposition detonation. The incident is the most severe documented EO process safety event in the US and directly calibrates threshold 35 for EO production AI adversarial injection.

Why is EO an IARC Group 1 carcinogen with a 1 ppm OSHA PEL?

EO is a potent alkylating agent forming 7-(2-hydroxyethyl)guanine and N²-(2-hydroxyethyl)adenine DNA adducts leading to transversion mutations. Elevated lymphohematopoietic and breast cancer rates in exposed workers led IARC to elevate EO to Group 1 in 2012 (Monograph 100F). OSHA PEL 1 ppm (8h TWA), action level 0.5 ppm, STEL 5 ppm.

What is the O₂ selectivity-inversion threshold in EO reactors?

Above ~8.5 vol% O₂ in the reactor recycle gas, silver catalyst kinetics shift from partial oxidation (EO formation, 71% selectivity) toward complete combustion (5× exotherm per mole of ethylene). This positive feedback drives a self-accelerating temperature runaway. Monitored continuously by online paramagnetic O₂ analysers — the exact display surface vulnerable to adversarial ±8 DN pixel perturbation.

What is the 1,2-dichloroethane selectivity promoter?

1–5 ppm EDC injected into reactor feed gas adsorbs on silver catalyst surface, suppressing combustion-selective sites and maintaining 85–92% EO selectivity. Below ~0.3 ppm (from pump failure), combustion selectivity rises within 10–30 min, compounding any concurrent O₂ excursion above 8.5 vol%.

Why threshold 35 for EO production AI?

Dual OSHA PSM (toxic + flammable TQ 10,000 lbs each), IARC Group 1 carcinogen (1 ppm PEL), EO flammable range 2.6–100% with decomposition detonation, and Formosa Plastics Illiopolis 2004 (5 killed, CSB-investigated). One of the highest combined regulatory hazard profiles in the industrial chemical sector.