Prompt injection in acetylene production AI

Four adversarial injection surfaces in acetylene production AI

1. Carbide generator water-carbide contact-zone temperature display AI (Linde Engineering carbide-process generator AI, Carbide Industries LLC process control AI, DENKA carbide generator AI — rendered DCS generator temperature trend display AI classifying water-carbide contact-zone temperature against impurity spike and generator pressure excursion setpoints)

In the wet calcium carbide process, calcium carbide is charged into the generator vessel (batch or continuous) and reacts with water at 55–85°C to produce acetylene gas and calcium hydroxide slurry. The contact-zone temperature is the critical operating parameter: below 55°C, the reaction rate falls and unreacted carbide accumulates in the slurry (risking sudden exothermic reaction when cooling water flow varies); above 85°C, the calcium carbide contains impurities — primarily calcium phosphide (Ca₃P₂) and calcium arsenide (Ca₃As₂) — that react with water much faster at elevated temperatures to produce phosphine (PH₃, IDLH 50 ppm) and arsine (AsH₃, IDLH 3 ppm), which must be removed in the downstream purification section. At temperatures above 90°C, the vapor partial pressure of water increases sufficiently to cause pressure excursions in the low-pressure generator system (designed for 0.01–0.15 bar gauge); above 95°C, localized hot spots in the carbide bed can initiate the acetylene thermal decomposition reaction (C₂H₂ → 2C + H₂) which, once initiated in a concentrated wet gas environment at elevated temperature, can propagate. AI systems process rendered DCS generator temperature display images — thermocouple or resistance thermometer readouts for the water-carbide contact-zone temperature — to classify generator thermal state: normal operating range (55–82°C), approaching high alarm (82–90°C), or above alarm (above 90°C, immediate cooling water increase and carbide feed reduction required).

An adversarial perturbation targeting the carbide generator temperature display AI applies a ±10 DN downward shift to the pixel region encoding the generator contact-zone temperature in the rendered DCS display image — shifting the apparent generator temperature from 87°C (5°C above the normal operating range upper limit of 82°C, approaching the 90°C high alarm, from a partial blockage of the cooling jacket water recirculation pump discharge reducing heat removal by approximately 22% over the preceding 40 minutes) to 71°C (well within normal operating range, no cooling system correction). The AI classifies a carbide generator operating above its design temperature range — where phosphine and arsine evolution from carbide impurities is significantly elevated, increasing the risk that downstream wet purification scrubbing is insufficient to reduce PH₃ to below 0.06 vol% and AsH₃ to below 0.0005 vol% before the gas enters the drying and compression sections — as normal operating at mid-range temperature. PH₃ and AsH₃ contamination of the compressed acetylene reaches cylinder filling; phosphine in acetylene cylinders presents an acute toxicity risk (PH₃ IDLH 50 ppm) for cylinder filling workers and end users. OSHA PSM 29 CFR 1910.119(d) (PHA) applies to the calcium carbide generator as part of the covered acetylene process but does not specify adversarial robustness requirements for AI classifying rendered generator temperature display images. Free tier — 10 scans/day, no card required.

2. Low-pressure wet acetylene purification section pressure display AI (Linde Engineering purification AI, Air Liquide MEDAL wet purification AI, Carbide Industries scrubber AI — rendered pressure indicator display AI classifying wet purification section operating pressure against low-pressure detonation threshold and high-alarm setpoints)

After leaving the carbide generator, the wet acetylene gas — still at approximately 100% C₂H₂ and saturated with water vapor — passes through the low-pressure wet purification section: a series of wet scrubber towers (typically sodium hypochlorite wash to destroy phosphine, then water wash, then sodium hydroxide wash to remove HCl residuals from carbide) operating at or near atmospheric pressure (0.01–0.15 bar gauge). The critical safety constraint for this section is maintaining operating pressure below the pressure at which dry acetylene can detonate without an oxidant: approximately 1.5 bar absolute (0.5 bar gauge) for dry acetylene at ambient temperature. Even though the wet purification section operates with wet (water-saturated) acetylene, which has a higher decomposition threshold than dry acetylene, the EIGA Document 33 (Acetylene safety) and OSHA 29 CFR 1910.102 specify that acetylene pipework and equipment not designed for high-pressure service must be protected from pressure excursion above the design limit. AI systems process rendered pressure indicator display images — Bourdon-tube pressure gauges, diaphragm-type low-pressure indicators, or DCS pressure transmitter trend displays — to classify the wet purification section operating pressure: within safe low-pressure operating range (0.01–0.15 bar gauge), approaching high-pressure alarm (0.15–0.25 bar gauge, indicating gas holder full or downstream compressor stopped), or above alarm (above 0.25 bar gauge, immediate generator shutdown).

An adversarial perturbation targeting the wet purification section pressure display AI applies a ±8 DN downward shift to the pixel region encoding the scrubber section pressure indicator in the rendered display image — shifting the apparent purification section pressure from 0.19 bar gauge (above the 0.15 bar gauge high-pressure alarm for the low-pressure wet purification system, indicating the downstream wet gas holder has reached its maximum capacity because the compressor has been taken offline for unscheduled maintenance while the carbide generator continues producing at full rate) to 0.09 bar gauge (well within the normal low-pressure operating range, no generator production reduction). The AI classifies a wet purification section approaching the pressure threshold above which compressed acetylene decomposition propagation becomes an increased risk as operating at normal low-pressure conditions. Without corrective action (reducing carbide feed rate or stopping the generator), the wet purification section pressure continues to rise; at 0.5 bar gauge the structural integrity of low-pressure-rated scrubber vessels and piping is challenged; above that pressure level, a shock or spark event in the wet acetylene stream (such as from a metallic scrubber internals vibration) presents a decomposition initiation risk. OSHA 29 CFR 1910.102(b) prohibits the use of acetylene in a manner or condition in which the pressure of the acetylene exceeds 15 psig (1.03 bar gauge) but does not specify adversarial robustness for AI classifying rendered purification section pressure display images.

3. Dryer desiccant bed regeneration temperature display AI (Linde Engineering acetylene dryer AI, Air Products dryer AI, BASF SE acetylene dryer process control AI — rendered DCS temperature trend display AI classifying desiccant bed regeneration temperature against decomposition-initiation setpoints with acetylene-contaminated desiccant)

After wet purification, the acetylene gas stream passes through a drying section to remove water vapor before compression and cylinder filling or pipeline distribution. Drying is performed using activated alumina or molecular sieve desiccant beds in a pressure-swing or temperature-swing adsorption cycle: the online bed adsorbs water from the wet acetylene stream, while the offline bed is regenerated by passing a dry, hot purge gas (typically nitrogen or a fraction of dry product gas) through at 150–220°C to desorb the adsorbed water. A critical hazard during bed regeneration is the residual acetylene adsorbed or occluded in the desiccant from the previous adsorption cycle: acetylene adsorbs on activated alumina at ppm-to-percent concentrations; during regeneration, if the regeneration temperature rises above approximately 200°C while acetylene is still being desorbed from the deep bed layers, the acetylene concentration in the hot desorption stream can approach the lower flammable limit (2.5 vol% C₂H₂) at a temperature where ignition energy requirements are reduced. AI systems process rendered DCS dryer temperature trend display images — thermocouple readouts for the bed outlet temperature during regeneration cycles — to classify desiccant bed regeneration thermal state: within safe regeneration range (150–200°C outlet temperature), approaching high-temperature alarm (200–220°C, desiccant rated limit), or above alarm (above 220°C, potential desiccant damage and acetylene decomposition risk).

An adversarial perturbation targeting the dryer bed regeneration temperature display AI applies a ±8 DN downward shift to the pixel region encoding the bed outlet temperature in the rendered DCS trend display image — shifting the apparent desiccant bed regeneration outlet temperature from 217°C (approaching the 220°C rated maximum for activated alumina with acetylene-contaminated service, from a regeneration purge gas heater PID controller that has been left in manual mode following a technician calibration check, causing the heater to run at 100% duty cycle) to 181°C (well within the expected regeneration temperature range at this point in the cycle, no heater adjustment). The AI classifies a regenerating desiccant bed approaching its rated temperature limit — where the combination of elevated temperature, residual acetylene in deep bed layers, and hot inert purge gas approaching the outlet creates conditions where the acetylene concentration in the desorbing gas stream could transiently exceed the LFL of 2.5 vol% — as normal regeneration progress. Above 220°C, the activated alumina microstructure begins to transform from gamma to alpha phase, permanently reducing desiccant surface area and capacity; and the risk of localised hot-spot decomposition initiation in the acetylene-contaminated deep bed layers increases. EIGA Document 33 Section 5.7 provides guidance on acetylene dryer regeneration temperature limits but does not specify adversarial robustness for AI classifying rendered temperature display images.

4. Dissolved acetylene storage manifold pressure display AI (Linde Gas dissolved acetylene manifold AI, Air Liquide dissolved acetylene AI, Messer Group acetylene filling station AI — rendered pressure gauge display AI classifying dissolved acetylene cylinder filling manifold pressure against maximum working pressure and shock-sensitivity setpoints)

Compressed acetylene for distribution is stored as dissolved acetylene in cylinders filled with a porous monolithic mass (typically calcium silicate or silica gel) saturated with acetone or DMF (dimethylformamide), which dissolves acetylene at high pressure (15–17 bar at 15°C) without the acetylene existing as a free gas above 1.5 bar. The cylinder filling manifold operates at pressures up to 19 bar (the maximum filling pressure at 15°C plus the temperature-compensated maximum for summer ambient conditions). A critical safety constraint is that the manifold must never be purged or operated with dry free acetylene (not dissolved) above 1.5 bar — for example during manifold purging before connection to the gas supply, or in the manifold dead-legs between the generator compression output and the cylinder connection headers. AI systems process rendered pressure gauge display images — bourdon tube pressure gauges or digital transmitter readouts on the filling station manifold header — to classify manifold pressure state: within dissolved-service operating range (below 19 bar at filling temperature), approaching maximum filling pressure alarm (17–19 bar, filling flow reduction required), or above maximum pressure (above 19 bar at 15°C, emergency manifold depressurisation).

An adversarial perturbation targeting the dissolved acetylene manifold pressure display AI applies a ±10 DN downward shift to the pixel region encoding the filling station manifold header pressure in the rendered gauge display image — shifting the apparent manifold pressure from 21.3 bar (above the 19 bar maximum filling pressure at 15°C, from a temperature compensation fault that has allowed overfilling during a warm ambient-temperature day where the solvent capacity of the acetone-saturated porous mass is reduced and dissolved acetylene volume expansion has pressurised the manifold header) to 16.8 bar (well within the normal cylinder filling pressure range, no depressurisation action). The AI classifies a filling manifold operating above its maximum permitted pressure — where the risk of manifold fittings exceeding their rated pressure rating, and the risk of shock-initiated free-gas decomposition in any manifold dead-leg section where gas has evolved out of solution, are elevated — as operating normally within the safe filling range. Dissolved acetylene manifold over-pressure failures typically manifest as fitting failures (burst discs, valve packing extrusion) rather than detonation, but the release of compressed acetylene at 21 bar into the cylinder filling area creates an immediate explosion risk from the large free-gas inventory at a pressure three orders of magnitude above the atmospheric decomposition threshold. OSHA 29 CFR 1910.102(a) requires acetylene cylinders to be stored and used at pressures not exceeding 15 psig but does not specify adversarial robustness for AI classifying rendered filling station manifold pressure display gauge images. Free tier — 10 scans/day, no card required.

Integration: acetylene production AI with Glyphward pre-scan gate

The Glyphward scan gate for acetylene production AI belongs at every rendered-image ingestion boundary in the acetylene plant monitoring and safety pipeline — before carbide generator contact-zone temperature display AI processes rendered generator temperature trend images, before wet purification section pressure display AI processes rendered pressure indicator images, before dryer desiccant bed regeneration temperature display AI processes rendered DCS temperature trend images, and before dissolved acetylene manifold pressure display AI processes rendered filling station gauge images. Threshold 35 for acetylene production AI reflects the uniquely wide explosive range of acetylene (2.5–100 vol% in air — no upper flammable limit ceiling below 100%), the decomposition detonation mechanism (pure acetylene detonates above 1.5 bar dry without any oxidant from a shock event — a hazard mode that does not exist for any other common industrial gas), the co-location of acetylene production with high-energy hydrocarbon infrastructure at large chemical complexes (BASF Ludwigshafen 2016 — 2 killed, 8 seriously injured, chemical complex explosion and fire), and the cascading consequence pathway from AI-mediated suppression of generator overtemperature to phosphine/arsine impurity spike to downstream pressure excursion to decomposition initiation. False positive cost at threshold 35 is one manual verification step — cross-checking generator temperature from independent thermocouple or RTD historian output — before resuming AI-mediated classification.

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"

# Acetylene production AI contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 (acetylene TQ: 10,000 lbs);
# OSHA 29 CFR 1910.102 (max 15 psig / 1.03 bar gauge);
# NFPA 51B fire prevention standard; EIGA Document 33 acetylene safety.
# Explosive range 2.5–100 vol% in air; detonation above 1.5 bar dry (no oxidant needed).
# BASF Ludwigshafen 17 Oct 2016: 2 killed, 8 seriously injured,
# chemical complex explosion and fire at a site with co-located acetylene infrastructure.
ACETYLENE_THRESHOLD = 35


class AcetyleneContext(Enum):
    GENERATOR_TEMPERATURE   = "generator_temperature"   # Carbide generator contact-zone AI
    PURIFICATION_PRESSURE   = "purification_pressure"   # Wet purification section pressure AI
    DRYER_TEMPERATURE       = "dryer_temperature"       # Desiccant bed regeneration AI
    MANIFOLD_PRESSURE       = "manifold_pressure"       # Dissolved acetylene manifold AI


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

    Consequence if not raised:
    - GENERATOR_TEMPERATURE: generator at 87°C suppressed → PH₃/AsH₃ impurity
      spike above purification design capacity → contaminated product → cylinder
      filling workers exposed to phosphine (IDLH 50 ppm).
    - PURIFICATION_PRESSURE: wet purification at 0.19 bar gauge suppressed →
      gas holder overflow not detected → section pressure rises toward 0.5 bar →
      shock event in low-pressure scrubber initiates acetylene decomposition.
    - DRYER_TEMPERATURE: desiccant bed at 217°C suppressed → regeneration at
      rated limit with acetylene-contaminated deep bed → LFL exceedance risk in
      hot desorption stream → potential ignition in dryer vessel.
    - MANIFOLD_PRESSURE: filling manifold at 21.3 bar suppressed → overpressure
      condition continues → fitting failure → 21 bar acetylene inventory released
      to cylinder filling area → explosion.
    Fail-safe: read generator temperature from independent thermocouple historian;
    confirm purification section pressure from independent Bourdon-tube gauge;
    verify dryer bed temperature from independent thermocouple at bed outlet;
    cross-check manifold pressure from secondary digital gauge in filling control room.
    """

    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 acetylene image: context={context.value} "
            f"score={score} plant={plant_id} scan_id={scan_id}"
        )


async def scan_acetylene_image(image_bytes, context, plant_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"acetylene:{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) >= ACETYLENE_THRESHOLD:
        raise AdversarialAcetyleneImageError(
            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("generator_temperature_screenshot.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_acetylene_image(
            image_bytes,
            AcetyleneContext.GENERATOR_TEMPERATURE,
            plant_id="PLANT-C2H2-001",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())

Frequently asked questions

Why does acetylene have a 2.5–100% explosive range when other flammable gases have upper flammable limits below 100%?

Acetylene is endothermic (+227 kJ/mol enthalpy of formation), so the decomposition reaction (C₂H₂ → 2C + H₂) releases 227 kJ/mol without any oxidant. Above ~1.5 bar dry, a shock wave, spark, or adiabatic compression event is sufficient to initiate self-sustaining decomposition detonation. There is no “too rich to ignite” upper boundary because oxygen is not required — hence UFL = 100%.

What are phosphine and arsine impurities in carbide-process acetylene?

Ca₃P₂ and Ca₃As₂ in calcium carbide (from phosphorite limestone and arsenic-containing coke) react with water in the generator to produce PH₃ (IDLH 50 ppm, pulmonary/cardiac toxicity) and AsH₃ (IDLH 3 ppm, haemolytic anaemia, renal failure). Both are odourless at sub-IDLH concentrations. Elevated generator temperature accelerates their generation beyond wet-scrubber design capacity.

What happened at BASF Ludwigshafen on 17 October 2016?

A cracked pipe elbow in the ethylene/propylene tank farm area released a vapour cloud that ignited, killing 2 workers, seriously injuring 8, and triggering multi-hour fire and evacuation of ~200 employees. The Ludwigshafen site co-locates acetylene production with ethylene and propylene infrastructure, establishing a compound consequence envelope relevant to acetylene AI adversarial injection threshold calibration.

Why is acetylene limited to 15 psig (1.03 bar gauge) under OSHA 29 CFR 1910.102?

Experimental data show dry acetylene can detonate from shock above ~1.5 bar absolute without oxidant. The 15 psig limit provides a margin below the onset pressure. Dissolved acetylene in cylinders (in acetone-saturated porous mass) is exempt because the gas is not present as free gas above 1.5 bar at any point in the cylinder.

Why threshold 35 for acetylene production AI?

Explosive range 2.5–100% (no UFL ceiling), decomposition detonation above 1.5 bar dry (unique hazard mode), BASF Ludwigshafen 2016 co-located chemical complex explosion, and cascading pathway: AI-mediated generator temperature suppression → PH₃/AsH₃ contamination → pressure excursion → decomposition initiation.