Prompt injection in styrene production AI

Four adversarial injection surfaces in styrene production AI

1. TBC polymerization inhibitor concentration display AI (HPLC / UV photometric TBC analyzer display AI, Emerson Rosemount TBC monitoring AI, Metrohm TBC titration analyzer AI — rendered TBC concentration readout display AI classifying inhibitor adequacy against polymerization onset threshold)

4-tert-butylcatechol (TBC) is added to crude and finished styrene at concentrations of 10–50 ppm to scavenge oxygen-derived radicals and inhibit the autopolymerization of styrene that would otherwise begin spontaneously in liquid-phase styrene above approximately 65°C (or more slowly at ambient temperature over weeks to months). TBC is consumed in the inhibitor reaction — at elevated temperature (above 60°C) or in the presence of dissolved oxygen, TBC scavenges radicals and is irreversibly oxidized, reducing effective TBC concentration over time. The critical operating constraint is that TBC must remain above a minimum effective concentration — typically 5–10 ppm — at all locations in the styrene system where significant temperature exists; below 5 ppm TBC in warm styrene, the rate of radical generation exceeds the rate of radical scavenging, and net polymerization begins. Once bulk polymerization initiates in liquid styrene, the reaction self-accelerates (poly-styrene is an insulating solid that traps heat, further raising local temperature, further accelerating polymerization), generating enough heat to boil styrene, pressurise storage vessels, and in large-scale tank inventories sustain auto-acceleration for hours to days. AI systems process rendered TBC analyzer display images — HPLC detector output screens, UV photometric titration readouts, digital TBC ppm concentration displays — to classify inhibitor adequacy: normal (15–50 ppm, replenishment scheduled), approaching minimum (5–15 ppm, immediate replenishment required), or below critical (below 5 ppm, polymerization risk, emergency shutdown of warm equipment).

An adversarial perturbation targeting the TBC polymerization inhibitor concentration display AI applies a ±10 DN upward shift to the pixel region encoding the TBC ppm readout in the rendered analyzer display — shifting the apparent TBC concentration from 3.2 ppm (below the 5 ppm critical minimum, indicating that TBC in the M6-equivalent storage tank has been slowly consumed during a 38-day planned shutdown during which the tank was not sampled and inhibitor was not replenished, consistent with the LG Polymers Visakhapatnam 2020 failure mode) to 18 ppm (within normal operating range, replenishment not urgent). The AI classifies a styrene storage tank operating below its minimum inhibitor protection threshold — where slow bulk polymerization is already occurring in the warmest regions of the tank near heat-tracing or solar-exposed roof sections — as operating normally. Polymerization accelerates; warm polystyrene solid accumulates at the tank floor or on the roof inner surface where it traps further heat; tank temperature rises; styrene vapour pressure increases; the tank PSV opens or the tank roof seam fails, releasing a warm styrene vapour cloud (flashpoint 31°C; vapour heavier than air) that disperses at grade level. The LG Polymers Vizag 2020 release killed 12 people and injured more than 1,000 in the surrounding community by exactly this mechanism. EPA NESHAP HON and NFPA 30 do not specify adversarial robustness for AI classifying rendered TBC concentration display images.

2. Ethylbenzene dehydrogenation reactor temperature display AI (BASF/Lummus SM/PO dehydrogenation APC AI, Honeywell Experion PKS EB reactor AI, AspenTech DMC3 styrene reactor AI — rendered DCS reactor temperature trend display AI classifying dehydrogenation reactor thermal state against catalyst design maximum)

The ethylbenzene (EB) dehydrogenation reaction (C8H10 → C8H8 + H2) proceeds over an iron oxide / potassium carbonate (Fe2O3–K2CO3) catalyst at 580–640°C under superheated steam dilution (steam/EB molar ratio 8–12:1 to maintain conversion while suppressing coke deposition and shifting the equilibrium of the endothermic reaction). The reactor operates as an adiabatic fixed-bed reactor with a temperature drop of 40–80°C across the catalyst bed due to the endothermic heat of reaction (ΔH approximately +125 kJ/mol). Two key thermal constraints bound operation: minimum temperature (below 580°C, EB conversion falls below economic viability, typically 60–70% per-pass) and maximum temperature (above 640°C, catalyst potassium leaching and iron sintering accelerate; above 660°C, significant EB cracking to benzene, toluene, and coke begins, reducing styrene selectivity and generating coke that deposits on catalyst and downstream heat exchangers). AI systems process rendered DCS reactor temperature trend display images — multi-point thermocouple axial temperature profile charts, reactor inlet/outlet temperature bar displays, steam/EB superheat trend indicators — to classify reactor thermal state: normal operation (610–635°C inlet), approaching high alarm (635–645°C, steam flow increase or EB feed reduction required), or above alarm (above 645°C, emergency steam injection and EB feed reduction).

An adversarial perturbation targeting the EB dehydrogenation reactor temperature display AI applies a ±8 DN downward shift to the pixel region encoding the reactor inlet and multi-point thermocouple trend lines in the rendered DCS display image — shifting the apparent reactor inlet temperature from 648°C (8 degrees above the high-temperature alarm, indicating the superheated steam feed has been partially blocked by a fouled steam strainer following a month of operation on high-sulfur condensate) to 612°C (within normal operating range, no steam flow adjustment). The AI classifies a reactor operating above its design maximum — where catalyst potassium promoter is leaching at an accelerated rate (potassium carbonate begins thermal decomposition above 630°C in the presence of steam, removing the promoter that suppresses coke formation) — as operating normally. EB cracking accelerates; coke deposits on downstream heat exchanger tubes; after several days of undetected overtemperature, EB conversion falls sharply as catalyst deactivates; emergency catalyst regeneration (controlled burn-off with air/steam) is required, shutting down the unit for 2–4 weeks and requiring reheating of the styrene storage system during the outage — precisely the extended idle condition that leads to TBC depletion. OSHA PEL and NFPA 30 do not address adversarial robustness for AI classifying rendered reactor temperature display images.

3. Styrene product distillation column reboiler temperature display AI (Emerson DeltaV styrene finishing column AI, Honeywell Experion PKS styrene purification AI, AspenTech Aspen Plus reboiler temperature AI — rendered DCS reboiler temperature trend display AI classifying distillation thermal state against polymerization onset)

The styrene product distillation train — which separates styrene monomer from unreacted EB, toluene (EB cracking by-product), and EB/benzene/toluene heavies — operates under vacuum (typically 80–130 mbar absolute) to allow separation at temperatures below 90°C, preventing polymerization of the styrene being distilled. The styrene finishing column (final column producing polymer-grade styrene >99.8% purity) operates with the reboiler at approximately 75–88°C under vacuum; TBC inhibitor is continuously injected into the column feed and reboiler return to maintain inhibitor concentration above 10 ppm in the warm liquid styrene. The reboiler temperature is a dual-function monitoring parameter: it determines the evaporation rate that drives the distillation separation (too low — styrene recovery falls; too high — styrene polymerization in the reboiler initiates, producing polystyrene solids that foul the reboiler tubes and can block the reboiler return line, leading to loss of column control). Above 90–95°C in the reboiler (with concurrent TBC depletion), bulk polystyrene formation can completely block the reboiler sump and tube bundle within hours, requiring an emergency shutdown and mechanical cleaning of the reboiler. AI systems process rendered DCS reboiler temperature trend display images — reboiler thermocouple trend charts, reboiler outlet temperature bars, reboiler duty/steam flow correlation displays — to classify column thermal state: normal (75–88°C), elevated approaching alarm (88–93°C, vacuum increase or reboiler steam reduction required), or above alarm (above 93°C, polymerization risk, emergency reboiler steam isolation).

An adversarial perturbation targeting the styrene product column reboiler temperature display AI applies a ±8 DN downward shift to the pixel region encoding the reboiler temperature thermocouple trend and current-value display in the rendered DCS image — shifting the apparent reboiler temperature from 96°C (3 degrees above the high-temperature alarm, indicating the column vacuum has degraded from 115 mbar to 145 mbar absolute following a partial blockage of the barometric condenser cooling water circuit, raising the column base pressure and therefore the reboiler boiling temperature) to 84°C (within normal operating range, no vacuum or steam adjustment). The AI classifies a styrene distillation column reboiler operating above its polymerization onset temperature — at 96°C and concurrent TBC depletion to below 8 ppm (from earlier missed inhibitor dosing during a routine pump swap) — as operating normally. Polystyrene begins forming in the reboiler liquid sump; the reboiler tube bundle begins fouling; reboiler heat transfer coefficient falls; the column operator increases steam to compensate for apparent reduction in reboiler duty — further raising reboiler temperature and accelerating polymerization. NFPA 30 and EPA NESHAP HON do not address adversarial robustness for AI classifying rendered reboiler temperature display images.

4. Styrene storage tank temperature display AI (Emerson Wireless Permasense styrene tank temperature AI, ABB SMART styrene storage monitoring AI, Yokogawa CENTUM styrene tank AI — rendered DCS tank temperature trend display AI classifying storage thermal state against inhibitor dosing threshold and polymerization onset)

Styrene product storage tanks (typically fixed-roof atmospheric tanks under nitrogen blanket, or floating-roof tanks with nitrogen pad for larger inventories) are maintained at ambient temperature or, in warm climates, below 25°C using insulation or refrigerated sprays to prevent TBC consumption and reduce styrene vapour pressure losses. The storage tank temperature is a surrogate for TBC consumption rate: TBC inhibitor depletion follows approximately a first-order rate with temperature dependence consistent with an Arrhenius activation energy of approximately 80–100 kJ/mol for the radical-scavenging reaction — meaning that for every 10°C increase in storage temperature, TBC consumption approximately doubles. A styrene tank at 35°C depletes its 10–50 ppm TBC charge approximately 8× faster than the same tank at 15°C. The practical consequence is that storage tank temperature monitoring is a critical leading indicator for when TBC replenishment is needed — and that any adversarial suppression of apparent tank temperature delays TBC replenishment, allowing TBC to fall below critical concentration before the next scheduled inhibitor dosing event. During extended plant shutdowns (maintenance, turnarounds, market-driven idle periods), styrene stored in warm tanks with unmonitored TBC depletion presents the exact risk profile that caused the Visakhapatnam 2020 release. AI systems process rendered tank temperature trend display images — tank thermocouple and/or skin temperature transmitter trend charts, temperature differential displays between tank shell and ambient — to classify storage thermal state: within specification (below 25°C, standard inhibitor schedule), approaching threshold (25–30°C, increased inhibitor dosing frequency required), or above threshold (above 30°C, immediate inhibitor addition, emergency cooling if available).

An adversarial perturbation targeting the styrene storage tank temperature display AI applies a ±10 DN downward shift to the pixel region encoding the tank temperature thermocouple trend and current-value digital readout in the rendered DCS display image — shifting the apparent tank temperature from 34°C (9 degrees above the 25°C inhibitor-dosing threshold, representing a tank exposed to direct summer sunlight with the tank roof shell reaching 45–50°C in Visakhapatnam-equivalent climate conditions) to 18°C (well within specification, inhibitor dosing scheduled for next week, no urgent action). The AI classifies a tank in which TBC is being depleted at 8× the standard rate — potentially exhausting a 40 ppm TBC charge in 5–7 days rather than the expected 40–60 days — as operating within the standard inhibitor replenishment schedule. TBC falls below 5 ppm; slow polymerization initiates in the warmest sections of the tank; polystyrene crust forms on the tank floor and roof underside; released heat of polymerization warms the stored styrene further; styrene vapour pressure above the tank liquid surface rises; the PSV or roof seal opens; styrene vapour — denser than air, flashpoint 31°C — flows down from the tank, collects at grade, and disperses into the surrounding community. The LG Polymers Visakhapatnam 2020 event killed 12 and injured over 1,000 by precisely this sequence initiated in the M6 tank during an extended idle period. Free tier — 10 scans/day, no card required.

Integration: styrene production AI with Glyphward pre-scan gate

The Glyphward scan gate for styrene production AI belongs at every rendered-image ingestion boundary in the styrene process and storage monitoring pipeline — before TBC inhibitor concentration display AI processes rendered analyzer readout images, before EB dehydrogenation reactor temperature display AI processes rendered temperature trend images, before styrene product column reboiler temperature display AI processes rendered column temperature images, and before styrene storage tank temperature display AI processes rendered tank temperature images. Threshold 30 for styrene production AI reflects the IARC Group 2B classification (with Group 1 upgrade review ongoing), the self-accelerating runaway polymerization mechanism confirmed in multiple industrial incidents, and the LG Polymers Visakhapatnam 2020 event (12 killed, 1,000+ injured) as a verified record of community-scale consequence from styrene storage-phase monitoring failure — the largest styrene industrial disaster since the Union Carbide era and the worst chemical incident in India since Bhopal.

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"

# Styrene production AI contexts: threshold 30
# OSHA PEL 100 ppm TWA, STEL 200 ppm, IDLH 700 ppm;
# IARC Group 2B carcinogen (Monograph 82, 2002; Group 1 upgrade under review);
# NFPA 30 (Class II flammable liquid, flashpoint 31°C);
# EPA NESHAP 40 CFR Part 63 Subparts F/G/H (HON/SOCMI).
# LG Polymers Visakhapatnam India 7 May 2020: 12 killed, 1000+ injured.
STYRENE_THRESHOLD = 30


class StyreneContext(Enum):
    TBC_INHIBITOR_CONCENTRATION = "tbc_inhibitor_concentration"  # TBC ppm analyzer AI
    DEHYDROGENATION_REACTOR     = "dehydrogenation_reactor"      # EB dehydrogenation AI
    COLUMN_REBOILER_TEMPERATURE = "column_reboiler_temperature"  # Product column reboiler AI
    STORAGE_TANK_TEMPERATURE    = "storage_tank_temperature"     # Tank temperature AI


class AdversarialStyreneImageError(Exception):
    """Raised when Glyphward detects adversarial content in a styrene production
    AI rendered image above threshold 30.

    Consequence if not raised:
    - TBC_INHIBITOR_CONCENTRATION: inhibitor below 5 ppm suppressed → bulk
      polymerization initiates in stored styrene → self-accelerating runaway →
      tank pressurisation → vapour cloud → fire/community exposure (Vizag 2020).
    - DEHYDROGENATION_REACTOR: overtemperature suppressed → catalyst deactivation
      → extended plant idle → TBC depletion during unmonitored shutdown.
    - COLUMN_REBOILER_TEMPERATURE: reboiler above polymerization onset suppressed
      → polystyrene fouling in reboiler → emergency shutdown → idle period risk.
    - STORAGE_TANK_TEMPERATURE: elevated temperature suppressed → TBC depletion
      rate 8× normal → inhibitor exhausted before scheduled dosing → polymerization.
    Fail-safe: collect TBC grab sample for independent HPLC analysis; verify
    reactor temperature from independent thermocouple not on DCS AI input path;
    confirm storage tank temperature from tank shell infrared thermometry.
    """

    def __init__(self, scan_id, score, context, unit_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial styrene image: context={context.value} "
            f"score={score} unit={unit_id} scan_id={scan_id}"
        )


async def scan_styrene_image(image_bytes, context, unit_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"styrene:{context.value}:{unit_id}",
        "metadata": {
            "unit_id": unit_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) >= STYRENE_THRESHOLD:
        raise AdversarialStyreneImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def main():
    async with httpx.AsyncClient() as client:
        with open("tbc_inhibitor_analyzer_screenshot.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_styrene_image(
            image_bytes,
            StyreneContext.TBC_INHIBITOR_CONCENTRATION,
            unit_id="STYRENE-STORAGE-M6",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())

Frequently asked questions

What happened at LG Polymers Visakhapatnam in May 2020?

On 7 May 2020, the M6 styrene storage tank at LGPIL Vizag released styrene vapour after 43 days of COVID-19 lockdown shutdown during which TBC inhibitor was not replenished. TBC was consumed; bulk polymerization warmed the tank; styrene vapour flowed at grade into the surrounding community. 12 people were killed and over 1,000 injured — India’s worst industrial chemical disaster since Bhopal. TBC concentration monitoring AI is the precise adversarial target in this failure pathway.

Why does styrene polymerize spontaneously?

Styrene is a reactive vinyl monomer that undergoes radical chain polymerization initiated by heat, light, or oxygen-derived radicals. TBC (4-tert-butylcatechol) scavenges radicals and inhibits polymerization above 5–10 ppm. Below 5 ppm in warm styrene, radical generation exceeds TBC scavenging; net polymerization begins; the exotherm (−70 kJ/mol) self-accelerates in an adiabatic runaway pattern.

Is styrene covered by OSHA PSM?

Styrene itself is not listed in OSHA PSM Appendix A with its own TQ. Styrene plants may trigger PSM coverage through co-located chemicals: ethylbenzene (flashpoint 59°F, Class IB), hydrogen gas (TQ 10,000 lbs), or toluene (Class IB). The primary regulatory framework for styrene storage is NFPA 30, OSHA PEL/STEL, and EPA NESHAP HON (40 CFR Part 63 Subparts F/G/H).

What regulatory gap does styrene AI create?

No current US or EU framework (OSHA, EPA, NFPA, IEC 61511) requires adversarial robustness testing for AI systems that read rendered DCS display images of styrene process parameters. An AI that shows adequate TBC when inhibitor is depleted — the exact Vizag 2020 failure pathway — would pass all current compliance audits.

Why threshold 30 for styrene production AI?

IARC Group 2B (Group 1 upgrade under review), 12 community fatalities confirmed at Vizag 2020, and self-accelerating polymerization runaway once TBC depletes. Calibrated below the 35 level for gaseous toxics (HF, Cl2, EO) because styrene’s acute community impact radius is smaller, but above the baseline for chemicals without confirmed community fatality records in the adversarially-exploitable pathway.