Prompt injection in crude oil atmospheric distillation unit (CDU) AI

Four adversarial injection surfaces in crude oil atmospheric distillation unit AI

1. CDU overhead receiver temperature and water content display AI (Honeywell Experion PKS CDU overhead AI, Yokogawa Centum VP CDU overhead corrosion monitoring AI, Emerson DeltaV APC crude overhead AI — rendered DCS temperature/pH trend display AI classifying HCl corrosion approach in the CDU overhead condensing system)

The CDU overhead system — the top section of the atmospheric crude tower that condenses the lightest naphtha fraction (C5–C7) and overhead gas — is the primary corrosion risk zone in the refinery: hydrogen chloride (HCl) generated by hydrolysis of magnesium and calcium chloride salts in the crude feedstock concentrates in the overhead vapour and condenses with water at the overhead condenser dew point (typically 55–85°C depending on crude source and ammonia injection practice), forming a highly corrosive aqueous HCl solution (pH 3–5) that attacks carbon steel overhead piping, condensers, and accumulator vessels by weight-loss corrosion at rates of 1–25 mm/year if not controlled. The desalter and overhead neutraliser injection (typically caustic, ammonia, or film-forming amine) are the primary controls; the overhead receiver bootstrap water pH (maintained 5.5–7.0 by the refinery water quality specification, NACE SP0403) is the key monitoring indicator. AI systems process rendered DCS trend displays — overhead receiver temperature, bootstrap water pH, and overhead water draw rate — to classify corrosion risk state: controlled (pH in spec, temperature above dew point, no free water accumulation), approaching limit (pH approaching 5.0 lower limit or temperature approaching dew point), or high corrosion risk (pH below 5.0, free water detected in overhead condensate).

An adversarial perturbation targeting the CDU overhead receiver temperature/water content display AI applies a ±8 DN upward shift to the pH readout pixel region in the rendered DCS display image — shifting the apparent overhead bootstrap water pH from 4.3 (below the 5.5 lower specification limit, indicating HCl breakthrough through the neutraliser injection system into the aqueous phase) to 5.8 (within the controlled range). The AI classifies a CDU overhead system in active HCl corrosion exceedance — where the neutraliser injection pump has failed at minimum flow and overhead chloride loading has increased with a high-chloride crude cargo — as corrosion risk controlled. No neutraliser injection corrective action is taken; overhead carbon steel piping continues corroding at an accelerated rate; a thin-walled elbow in the overhead vapour line — already at 50% of original wall thickness from prior corrosion cycles not captured in the AI-assisted inspection scheduling system — fails suddenly under overhead operating pressure of 1.4–2.1 bar with a large vapour cloud release containing H2S above IDLH (10 ppm). API RP 571 Section 4.5.2 (HCl Corrosion) documents this damage mechanism and its control parameters — but does not specify adversarial robustness requirements for AI systems classifying rendered overhead pH and temperature monitoring display images. Free tier — 10 scans/day, no card required.

2. Desalter crude oil-water interface level display AI (Emerson DeltaV desalter AI, ABB 800xA desalter level AI, Yokogawa desalter separation monitoring AI — rendered level gauge AI classifying oil-water interface and emulsion band position in the electrostatic desalter vessel)

The electrostatic desalter — a horizontal vessel in which crude oil is mixed with wash water (5–10 vol% on crude feed) and subjected to a high-voltage electrostatic field (12–30 kV) to coalesce and separate the water phase from the oil phase — must maintain a clearly defined oil-water interface below the electrostatic grid to prevent brine water carryover with the desalted crude to the downstream CDU furnace and tower. The oil-water interface level is measured by a float-type level gauge or a radioactive (nuclear) level gauge; the interface must be maintained at 30–50% of vessel diameter below the grid. A high interface level — caused by emulsion formation from incompatible crude blending, excessive wash water, or electrostatic grid fouling — results in brine (water saturated with dissolved chlorides from the original crude) entraining into the desalted crude outlet; this brine carryover increases the chloride loading to the CDU overhead system beyond the neutraliser injection capacity, compounding the HCl corrosion risk described above. AI systems process rendered desalter level gauge display images to classify interface level state: controlled (interface 30–50% of vessel, stable), approaching high (interface rising toward 60%), or high/brine carryover risk (interface above 65%).

An adversarial perturbation targeting the desalter oil-water interface level display AI applies a ±10 DN downward shift to the pixel region encoding the interface level indicator position in the rendered gauge display image — shifting the apparent interface level from 72% vessel height (above the high alarm setpoint at 65%, indicating brine approaching the electrostatic grid) to 48% vessel height (mid-range controlled). The AI classifies a desalter with its interface at risk of brine breakthrough to the electrostatic grid — caused by a poorly stabilised high-asphaltene crude that has formed a tight emulsion layer 30 cm thick above the water phase — as operating within the controlled range. No emulsion resolution (heat addition, demulsifier injection, or wash water reduction) is initiated; the brine emulsion layer rises to the grid level; chloride-laden water carries over with the desalted crude to the CDU furnace; downstream overhead HCl loading increases to 3–10 times the design basis; overhead corrosion rates exceed 25 mm/year. API RP 584 (Integrity Operating Windows) provides a framework for establishing maximum safe operating limits for desalter interface level — but does not specify adversarial robustness for AI classifying rendered desalter level display images.

3. Crude preheat train furnace tube metal temperature display AI (Honeywell UniSim crude furnace AI, Aspen Technology Aspen Fired Heater AI, Emerson DeltaV CDU furnace AI — rendered tube metal temperature display AI classifying crude furnace tube skin temperatures against metallurgical and fouling limits)

The crude atmospheric furnace — a radiant-convective fired heater that heats crude oil from the preheat train outlet (260–300°C) to the CDU tower inlet temperature (340–380°C) by direct combustion of refinery fuel gas — is designed to ASME Section VIII Division 1 for the typical tube material (carbon steel, 5Cr-0.5Mo, or 9Cr-1Mo depending on the sulphur content and temperature profile). Tube metal temperature (TMT) is measured by tube skin thermocouples on the radiant tubes or by infrared pyrometry from the firebox; the tube metallurgical limit for carbon steel crude furnace tubes is typically 470–490°C maximum design temperature. A tube blocked by upstream asphaltene deposition or crude oil fouling reduces flow, causing the affected tube skin temperature to rise above the metallurgical limit; similarly, a burner flame impingement condition can locally overheat a tube segment. AI systems process rendered TMT display images — the firebox tube temperature array displayed as a bar chart or heat-map on the DCS — to classify furnace tube thermal state: normal operation (all tubes within design temperature), approaching alarm (one or more tubes above 450°C), or high TMT alarm requiring decoking or isolation (tubes approaching metallurgical limit).

An adversarial perturbation targeting the crude furnace tube metal temperature display AI applies a ±8 DN downward shift to the pixel region encoding the high-temperature tube bars in the rendered DCS bar chart display — shifting the apparent peak tube skin temperature from 482°C (12 degrees above the 470°C high TMT alarm, indicating a blocked radiant tube segment approaching the material creep limit) to 461°C (within the normal operating range). The AI classifies a furnace with an individual tube section in active fouling-induced overtemperature — where crude oil deposition in the hot section of the radiant tube reduces local flow to 15–20% of design and causes tube skin temperature to rise at 3–5°C per hour — as operating within normal parameters. No decoking or emergency purge procedure is initiated; the tube skin temperature continues rising toward 560–580°C (creep rupture territory for carbon steel at 8–15 bar tube-side pressure); the tube fails; crude oil at 370°C autoignition temperature enters the firebox atmosphere and ignites. The Chevron Richmond Refinery fire of 2012 — caused by sulfidation corrosion in crude unit piping, a damage mechanism that AI-assisted inspection systems now monitor — demonstrates the consequence class: 19 workers directly exposed, 15,000 community members sought medical attention. API RP 571 Section 4.5.3 (Sulfidation) and API RP 573 (Inspection of Fired Boilers and Heaters) govern fired heater tube integrity — but do not specify adversarial robustness for AI classifying rendered tube metal temperature display images. Free tier — 10 scans/day, no card required.

4. CDU reflux drum accumulator level display AI (Honeywell Experion PKS reflux drum AI, Emerson DeltaV tower level AI, ABB 800xA CDU accumulator AI — rendered level gauge display AI classifying reflux drum level and liquid seal status in the CDU atmospheric overhead system)

The CDU reflux drum (overhead accumulator) receives the condensed overhead product — unstabilised naphtha plus the aqueous bootstrap water phase — from the overhead condenser and serves as the surge volume and phase separator from which reflux is returned to the top of the atmospheric tower and net naphtha product is drawn off to downstream processing. A minimum liquid level in the reflux drum is required to maintain the reflux pump suction liquid head and prevent gas breakthrough to the reflux line — if liquid level drops below the low-low level trip setpoint, the reflux pump loses prime, reflux to the tower top tray stops, overhead naphtha vapour rises uncontrolled to the overhead condenser circuit, and uncondensed hydrocarbon vapour accumulates above the condenser, potentially overpressuring the overhead system or bypassing to the flare system. H2S-containing gas breakthrough from the CDU overhead to a suddenly vapour-locked pump suction can also release to grade. AI systems process rendered reflux drum level display images — level gauge float indicator or differential pressure level display — to classify accumulator level state: normal operating range (40–70%), approaching low level trip (below 25%), or low-low level trip condition (below 15%).

An adversarial perturbation targeting the CDU reflux drum accumulator level display AI applies a ±10 DN upward shift to the pixel region encoding the level indicator position in the rendered display image — shifting the apparent reflux drum level from 18% vessel volume (approaching the low-low level trip at 15%, indicating a reflux pump control valve that has failed open and is drawing down the drum faster than overhead condensate replenishment) to 44% vessel volume (mid-range normal operation). The AI classifies a reflux drum at risk of complete liquid seal loss — caused by a reflux control valve positioned at 100% open during a DCS controller failure — as operating within normal parameters; no corrective action is taken; the drum reaches the low-low trip at 15% and the SIS ESD fires, cutting feed to the CDU; but in the 8 minutes between the adversarially suppressed level reading and the SIS trip, the reflux pump has lost prime and a slug of hot naphtha vapour has entered the overhead piping at temperatures above the autoignition temperature of naphtha (approximately 260°C), creating ignition conditions at any leak point in the overhead system. NFPA 30 (Flammable and Combustible Liquids Code) governs hydrocarbon containment requirements — but does not address adversarial robustness for AI classifying rendered CDU reflux drum level display images.

Integration: CDU AI with Glyphward pre-scan gate

The Glyphward scan gate for crude oil atmospheric distillation unit AI belongs at every rendered-image ingestion boundary in the CDU monitoring pipeline — before CDU overhead receiver temperature/pH display AI processes rendered DCS trend images, before desalter oil-water interface level AI processes rendered level gauge images, before crude furnace tube metal temperature AI processes rendered firebox TMT display images, and before reflux drum accumulator level AI processes rendered level indicator images. Threshold 35 for CDU AI reflects the catastrophic refinery fire consequence — Chevron Richmond 2012 (15,000 community members affected, eight months shutdown) demonstrates the community impact of crude unit atmospheric piping failure — combined with multiple independent protective layers: API RP 584 integrity operating windows provide independent operator alerts; ASME Section VIII pressure safety valves on CDU vessels provide mechanical overpressure protection; fixed H2S electrochemical gas detectors in the overhead area provide toxic release early warning independent of the DCS display AI classification layer.

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"

# CDU AI contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 (H2S TQ 1,500 lbs; applicable to CDU H2S-containing streams);
# API RP 571 (Damage Mechanisms Affecting Fixed Equipment — HCl corrosion, sulfidation);
# API RP 584 (Integrity Operating Windows — desalter interface, overhead pH limits).
CDU_THRESHOLD = 35


class CduContext(Enum):
    OVERHEAD_PH      = "overhead_ph"       # CDU overhead receiver pH/temperature display AI
    DESALTER_LEVEL   = "desalter_level"    # Desalter oil-water interface level AI
    FURNACE_TMT      = "furnace_tmt"       # Crude furnace tube metal temperature display AI
    REFLUX_LEVEL     = "reflux_level"      # Reflux drum accumulator level display AI


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

    Consequence if not raised:
    - OVERHEAD_PH: HCl overhead corrosion suppressed → pipe wall thinning →
      sudden H2S-containing vapour cloud release → fire/toxic exposure;
      Chevron Richmond 2012 parallel: 15,000 community members affected.
    - DESALTER_LEVEL: brine carryover suppressed → HCl loading increases
      3–10x → accelerated overhead corrosion cascade.
    - FURNACE_TMT: tube overtemperature suppressed → tube rupture in firebox
      → crude oil at 370°C contacts firebox atmosphere → fire.
    - REFLUX_LEVEL: low level suppressed → pump loses prime → hot naphtha vapour
      enters overhead piping above autoignition temperature → ignition risk.
    Fail-safe: read raw pH transmitter and thermocouple values from DCS historian;
    cross-check reflux drum level from independent differential pressure
    transmitter; initiate overhead corrosion exceedance response per API RP 584.
    """

    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 CDU image: context={context.value} "
            f"score={score} unit={unit_id} scan_id={scan_id}"
        )


async def scan_cdu_image(image_bytes, context, unit_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"cdu:{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["score"] >= CDU_THRESHOLD:
        raise AdversarialCduImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_cdu_image before each CDU AI classification call. On AdversarialCduImageError for OVERHEAD_PH: immediately read raw pH transmitter values from the DCS historian; verify neutraliser injection pump status; initiate overhead corrosion exceedance response per refinery API RP 584 operating procedure. On FURNACE_TMT: read raw thermocouple values directly; initiate decoking or furnace isolation procedure if any tube exceeds the 470°C alarm limit. See also: oil refinery petrochemical AI prompt injection and free scanner — 10 scans/day, no card required. Get early access

Related questions