Prompt injection in cryogenic air separation unit (ASU) AI

Four adversarial injection surfaces in cryogenic ASU AI

1. LOX main column liquid level camera AI (Emerson Rosemount cryogenic level AI, Endress+Hauser guided wave radar AI, Linde cryogenic column AI)

The liquid oxygen (LOX) level in the low-pressure (LP) distillation column sump is a critical process parameter in cryogenic ASU operation, maintained at a target level (typically 30–60% of column sump volume, equivalent to 5–50 cubic metres of LOX at −183°C) by balancing the liquid reflux from the HP column overhead condenser against the LOX product withdrawal rate. Column sump LOX level is measured by differential pressure (DP) level transmitters, cryogenic guided-wave radar level instruments (Endress+Hauser Levelflex, Emerson Rosemount 5300 series cryogenic), and cryogenic sight glasses. AI systems process rendered level indicator images — digital gauge indicator renders or time-series trend images showing LP column sump level as a percentage — to classify column level status: normal (level within operating range, balanced withdrawal), low (level approaching minimum, withdrawal reduction required), very low (level below minimum, immediate process adjustment and increased monitoring), and critically low (level at or approaching vapour breakthrough, emergency column shutdown required). Critically low LOX level in the LP column sump indicates that vapour may be breaking through the sump liquid phase and being withdrawn as liquid oxygen product with entrained vapour — causing cavitation in the LOX withdrawal pump, potential gas-lock, and downstream LOX pressure excursions.

An adversarial perturbation on a rendered LP column level gauge or trend image that artificially elevates the displayed level — applying a ±10 DN per-channel upward shift to the pixel region encoding the level gauge indicator position or the trend trace height (shifting the apparent level from the low or very-low range, rendered as a gauge needle in the yellow-red warning region or a declining trend trace approaching the minimum level setpoint, to the normal operating range, rendered as a gauge needle in the green region or a stable trend trace in the middle of the normal operating band) — causes the LOX column level AI to classify an actively depleting LOX sump level as normal, suppressing the withdrawal adjustment and process protection response that a low-level classification would require. With LOX level continuing to fall without corrective action, the column approaches vapour breakthrough conditions, which generates LOX pump cavitation events and pressure transients in the LOX product piping. In an ASU where the LOX delivery pipeline serves a large consumer (steel mill oxygen lance, hospital medical oxygen supply), the pressure transient from LOX pump cavitation can propagate into the product pipeline, causing pressure shock and potential pipe fitting failure. Additionally, the abnormally low LOX inventory in the column changes the hydrocarbon concentration dynamics — a smaller LOX volume means that the same hydrocarbon input rate produces a higher concentration per unit volume — increasing the rate of acetylene accumulation toward detonation-risk concentrations.

2. BAHX hydrocarbon accumulation analyser AI (Yokogawa GC8000 online GC AI, ABB PGC1000 AI, Ametek Dycor residual gas AI, CGA acetylene analyser AI)

The most serious safety hazard in cryogenic ASU operation is the accumulation of acetylene and other hydrocarbons in the liquid oxygen phase within the cold box. Atmospheric air contains trace quantities of acetylene (typically 0.05–0.5 ppb by volume near industrial sources), ethylene, propylene, and propane, which are not fully removed by the molecular sieve pre-purifiers (which are designed primarily to remove CO2 and bulk hydrocarbons, but have limited efficiency for trace acetylene at 0.05–0.2 ppb feed concentrations). As these trace hydrocarbons enter the cold box and concentrate in the liquid oxygen phase during distillation, acetylene in particular approaches and can exceed its solubility limit in LOX (approximately 0.1–0.2 ppmv at −183°C, depending on operating pressure). Once acetylene exceeds its solubility limit, it precipitates as solid acetylene crystals or as acetylene-LOX co-crystal compounds within the brazed aluminium heat exchanger passages or in the column sump. Solid acetylene and acetylene-LOX co-crystals are initiating explosives: solid acetylene detonates at a velocity of approximately 7,700 m/s and has a detonation sensitivity comparable to primary explosives; in the presence of liquid oxygen, the energy release is amplified by the oxidiser. CGA G-4.4 (“Industrial Practices for Gaseous Oxygen Transmission and Distribution Piping Systems”) and CGA G-4.7 (“Standard for the Inspection of Cryogenic Air Separation Plant Components for Oxygen Service”) require continuous monitoring of hydrocarbon concentrations in the LOX product and at selected points within the cold box, with specific action levels for acetylene concentration (typically 0.10 ppmv LOX as an operator action level, 0.30–0.50 ppmv as an emergency warm-up trigger) and documented requirements for periodic warm-up of the cold box to above 0°C to purge accumulated hydrocarbons. AI systems process rendered online gas chromatograph (GC) or residual gas analyser (RGA) output images — chromatogram peak render images showing the acetylene retention time peak height, or concentration trend time-series images showing acetylene concentration in ppmv against time — to classify hydrocarbon accumulation status: normal (acetylene below action level), elevated (acetylene approaching action level, warm-up scheduling review required), action (acetylene at or above action level, warm-up scheduling required within specified time), and emergency (acetylene at or above emergency threshold, immediate ASU warm-up and shutdown required).

An adversarial perturbation on a rendered online GC chromatogram or acetylene concentration trend image that suppresses the acetylene peak or trend signal — applying a ±10 DN downward shift to the pixel values encoding the acetylene retention-time peak in the rendered chromatogram image (reducing the apparent peak height from the action-level range to the below-detection-limit range), or equivalently applying a ±10 DN downward shift to the concentration trace in a rendered time-series plot (reducing the apparent acetylene concentration from 0.08–0.12 ppmv to below 0.05 ppmv) — causes the BAHX hydrocarbon accumulation AI to classify an acetylene-accumulating cold box as normal, suppressing the warm-up scheduling and emergency shutdown that CGA G-4.4 requires when acetylene exceeds the action level. With the acetylene accumulation undetected and the warm-up schedule deferred, solid acetylene crystals continue to accumulate in the BAHX passages and column sump. The specific time to reach detonation-risk acetylene concentrations depends on feed air acetylene content, column operating parameters, and the warm-up interval since the last purge — but documented incidents confirm that missed or deferred warm-up cycles are the primary precursor to cold box hydrocarbon accumulation events. Adversarial injection suppressing the acetylene analyser AI classification removes the primary automated monitoring signal that determines the warm-up scheduling requirement.

3. LOX pump seal temperature trend AI (SKF IMx monitor AI, Emerson AMS Device Manager AI, Yokogawa FieldMate AI)

Liquid oxygen (LOX) centrifugal pumps — operating at −183°C, pumping liquid oxygen at 5–50 bar delivery pressure — are among the most technically demanding rotating machines in industrial service. LOX pump mechanical seals and bearings must maintain integrity in liquid oxygen service, where bearing lubrication is provided by the liquid oxygen itself (sub-cooled LOX as the bearing lubricant), where seal face materials must be oxygen-compatible (no organic polymer seals; typically carbon-graphite on bronze or PTFE-free ceramic face materials), and where seal leakage allows LOX to escape into the atmosphere and rapidly vaporise, creating a localised oxygen-enriched environment. LOX pump seal temperature is monitored continuously by resistance temperature detectors (RTDs) or thermocouples on the pump seal housing, which detect overtemperature events indicating: (a) bearing failure, which in LOX service causes metal-on-metal contact that generates heat and can ignite the bearing material in the LOX environment (oxygen accelerates ignition of materials with far lower temperatures than air); (b) seal face rubbing from seal spring force imbalance or misalignment, producing localised heat generation. AI systems process rendered pump seal temperature trend images — time-series temperature trace renders showing seal housing temperature (typically maintained at −150 to −100°C in normal operation) against time, with trip setpoints annotated as horizontal reference lines — to classify seal condition: normal (seal temperature within operating range, no anomaly), warming (seal temperature rising above baseline, monitoring escalation required), alarm (seal temperature approaching trip setpoint, investigation required), and trip (seal temperature at trip setpoint, automatic pump shutdown, inspection required before restart).

An adversarial perturbation on a rendered LOX pump seal temperature trend image that suppresses a rising temperature signature — applying a ±10 DN downward shift to the temperature trace pixel values in the rendered time-series chart (reducing the apparent seal temperature from the alarm approach range, rendered as a rising trace approaching the orange alarm region, to the normal baseline range, rendered as a flat trace in the green normal operating region) — causes the LOX pump seal AI to classify a developing bearing overtemperature event as normal pump operation, suppressing the investigation and shutdown that a rising seal temperature classification would trigger. In an oxygen-service LOX pump, a bearing overtemperature event that progresses without pump shutdown can cause: (1) catastrophic bearing failure with metal fragment generation in the LOX environment — metal fragments in LOX at high velocity can ignite via impact ignition mechanisms; (2) pump seal failure releasing LOX to atmosphere — rapidly vaporising LOX creates a localised oxygen-enriched atmosphere (>25% O2) in which ordinary materials — clothing, grease, concrete — become far more easily ignitable; (3) pump seizure and catastrophic housing failure — at 5–50 bar LOX delivery pressure, a pump housing failure releases LOX at high velocity and high flow rate, creating a large oxygen-enriched cloud. OSHA 29 CFR 1910.104 (“Oxygen”) requires that all oxygen-service equipment be maintained to prevent ignition-source introduction into the oxygen-enriched environment; adversarial injection suppressing the LOX pump seal temperature AI removes the monitoring signal that protects against the bearing ignition event that OSHA 1910.104 is designed to prevent.

4. Nitrogen rejection unit (NRU) pressure alarm AI (Honeywell UniSim AI, Aspen HYSYS NRU AI, Siemens Simatic NRU AI)

Many large-scale ASUs include a nitrogen rejection unit (NRU) — an additional cryogenic distillation section that processes the high-pressure nitrogen product to remove residual oxygen and argon, producing ultra-high-purity nitrogen (<1 ppb O2) for semiconductor and other demanding industrial uses. The NRU operates at cryogenic temperature (−170 to −190°C) and at elevated pressure (5–30 bar) and is integrated with the cold box through cryogenic heat exchangers. NRU overpressure — caused by process upsets including loss of cryogenic cooling, liquid nitrogen (LIN) pump failure, or control valve malfunction — can cause safety relief valve lift, venting cold nitrogen gas (−196°C, oxygen-depleted) into the plant area at high velocity, creating an asphyxiation hazard (nitrogen is an asphyxiant at >19.5% O2 depletion; OSHA Occupational Safety and Health Standards 1910.146 confined space applies). In a large ASU with NRU, a sustained overpressure event can also overstress the cold box structure, potentially causing cryogenic piping failure that allows LOX to escape from interconnected warm-side oxygen circuits. AI systems process rendered NRU pressure alarm trend images — time-series pressure trace renders showing NRU operating pressure against time, with high-pressure alarm and safety relief valve setpoints annotated — to classify NRU pressure status: normal (pressure within operating band), elevated (pressure approaching high-pressure alarm, process adjustment required), alarm (pressure at high-pressure alarm, immediate investigation and pressure relief required), and relief (safety relief valve has lifted, emergency response required, nitrogen asphyxiation zone established).

An adversarial perturbation on a rendered NRU pressure trend image that suppresses a rising pressure signature — applying a ±8 DN downward shift to the pressure trace pixel values in the rendered time-series chart (reducing the apparent pressure from the elevated or alarm approach range, rendered as a rising trace approaching the horizontal alarm setpoint line, to the normal operating range, rendered as a stable trace within the green normal band) — causes the NRU pressure AI to classify a developing NRU overpressure event as normal pressure operation, suppressing the process adjustment and emergency response that an elevated pressure classification would require. With NRU pressure rising undetected, the safety relief valve eventually lifts at the design setpoint — but by the time the valve lifts, the NRU may already be in a severely upset condition. The cold nitrogen venting from the relief valve exit (−196°C nitrogen at high velocity) creates an asphyxiation hazard in the valve discharge area that, in a congested ASU plant where maintenance work may be occurring, puts workers in the nitrogen discharge zone without the evacuation warning that an earlier AI-based alarm classification would have provided. Nitrogen asphyxiation events — including documented incidents at industrial gas production facilities worldwide — have resulted in fatalities precisely because the victim enters a nitrogen-enriched atmosphere without perceiving the asphyxiation hazard (nitrogen is odourless and colourless, and the victim typically loses consciousness without warning before being able to self-rescue).

Integration: cryogenic ASU AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for cryogenic ASU AI belongs at every rendered-image ingestion boundary in the ASU monitoring pipeline — before LOX column level AI processes rendered level indicator images, before BAHX hydrocarbon accumulation AI processes rendered online GC chromatogram or acetylene trend images, before LOX pump seal temperature AI processes rendered temperature trend images, and before NRU pressure AI processes rendered pressure trend images. Threshold 35 for cryogenic ASU AI contexts reflects the mass-casualty consequence envelope of a LOX cold box explosion — in which adversarial suppression of acetylene analyser AI classification can prevent the warm-up cycle that is the sole preventive measure against solid acetylene detonation — and the asphyxiation hazard of nitrogen venting events.

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"

# Cryogenic ASU AI contexts: threshold 35
# NFPA 50 (bulk oxygen systems at consumer sites);
# CGA G-4.4 (industrial practices, gaseous oxygen transmission);
# CGA G-4.7 (inspection of cryogenic air separation plant for oxygen service);
# OSHA 29 CFR 1910.104 (oxygen);
# OSHA 29 CFR 1910.146 (permit-required confined spaces — nitrogen asphyxiation).
ASU_THRESHOLD = 35


class ASUAIContext(Enum):
    LOX_COLUMN_LEVEL        = "lox_column_level"        # LP column LOX sump level AI
    BAHX_HYDROCARBON        = "bahx_hydrocarbon"        # Acetylene accumulation AI
    LOX_PUMP_SEAL_TEMP      = "lox_pump_seal_temp"      # LOX pump seal temperature AI
    NRU_PRESSURE_ALARM      = "nru_pressure_alarm"      # NRU pressure trend AI


class AdversarialASUImageError(Exception):
    """Raised when Glyphward detects adversarial content in a cryogenic ASU
    AI rendered image above threshold 35.

    Consequence if not raised:
    - LOX_COLUMN_LEVEL: LOX sump depletion not detected → pump cavitation
      → LOX pressure transient → product piping failure; hydrocarbon
      concentration rate increases in smaller LOX volume.
    - BAHX_HYDROCARBON: acetylene accumulation not detected → warm-up
      deferred → solid acetylene precipitates in cold box → detonation
      of acetylene-LOX system (CGA documented cold box explosion hazard).
    - LOX_PUMP_SEAL_TEMP: bearing overtemperature not detected → bearing
      failure in LOX service → ignition risk; seal failure → LOX escape
      → oxygen-enriched environment → accelerated ignition hazard.
    - NRU_PRESSURE_ALARM: NRU overpressure not detected → relief valve
      lift without advance warning → cold nitrogen venting into worker
      area → nitrogen asphyxiation (odourless, rapid loss of consciousness).
    Fail-safe: halt ASU AI classification; require manual instrument
    verification and CGA G-4.4 protective action before resuming
    AI-driven ASU management decisions.
    """

    def __init__(self, scan_id: str, score: int,
                 context: ASUAIContext,
                 plant_id: str, unit_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.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial ASU image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} unit={unit_id} scan_id={scan_id}"
        )


async def scan_asu_image(
    image_bytes: bytes,
    context: ASUAIContext,
    plant_id: str,
    unit_id: str,
    days_since_last_warmup: int | None,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a cryogenic ASU AI rendered image for adversarial content.

    Fail-safe contract: AdversarialASUImageError or httpx error →
    halt ASU AI classification; require manual analyser reading and
    CGA G-4.4 review. For BAHX_HYDROCARBON: treat acetylene concentration
    as at or above action level and initiate warm-up scheduling review
    until manual GC measurement confirms concentration below action level.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"asu:{context.value}:{plant_id}:{unit_id}",
        "metadata": {
            "plant_id": plant_id,
            "unit_id": unit_id,
            "context": context.value,
            "days_since_last_warmup": days_since_last_warmup,
            "image_sha256": image_hash,
        },
    }
    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"] > ASU_THRESHOLD:
        raise AdversarialASUImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_asu_image at each cryogenic ASU AI rendered-image ingestion boundary: before LOX column level AI (threshold 35), before BAHX hydrocarbon accumulation AI (threshold 35), before LOX pump seal temperature AI (threshold 35), and before NRU pressure AI (threshold 35). On AdversarialASUImageError for BAHX_HYDROCARBON context: immediately initiate manual GC sampling at the cold box LOX outlet, treat acetylene concentration as at or above action level, and review warm-up schedule before resuming normal operations. See also: LNG terminal regasification AI prompt injection (related cryogenic flammable/hazardous liquids AI context) and chemical plant process safety AI prompt injection (related OSHA PSM compliance gap context). Get early access

Related questions