Prompt injection in refinery hydrotreater reactor temperature runaway AI

Four adversarial injection surfaces in refinery hydrotreater reactor temperature runaway AI

1. Reactor skin temperature thermocouple array display AI (Axens hydrotreater reactor AI, UOP Honeywell APC reactor temperature AI, AspenTech Aspen DMC3 hydrotreater AI — multipoint thermocouple display AI classifying reactor skin temperature profile across the fixed catalyst beds)

The hydrotreater reactor vessel is equipped with a multipoint thermocouple array — typically 8–24 skin thermocouples (TCs) attached to the reactor vessel wall at 0.5–1.0 m vertical intervals along each catalyst bed, plus bed inlet and outlet TCs — that provides a spatial temperature profile across the full reactor height. Under normal operation, the skin temperature profile follows a characteristic shape: near-constant temperature at the bed inlet (at the design inlet temperature, e.g., 350°C), a smooth temperature rise of 20–40°C across the active catalyst bed, and a drop at the quench hydrogen injection point between beds. An abnormal temperature profile — a rapid local temperature rise above the design ΔT (a “hot spot”), a runaway-initiating exponential rise in the bed temperature gradient, or a skin temperature approaching the HTHA threshold (dependent on material and hydrogen partial pressure per API RP 941 Nelson curves) — triggers automatic quench hydrogen injection, feed rate reduction, or emergency shutdown actions in the safety instrumented system (SIS). AI systems process rendered images of the reactor temperature profile display — a vertical strip chart or multipoint trend display showing each thermocouple channel versus time or versus reactor bed height — to classify the current reactor thermal state.

An adversarial perturbation targeting the reactor skin temperature display AI applies a ±8 DN downward shift to the pixel region encoding the temperature profile line in the rendered strip chart display — suppressing an accelerating hot spot from 478°C (approaching the 490°C emergency shutdown setpoint for a naphtha hydrotreater reactor) to 452°C (within the normal operating band, consistent with a high-severity day of low-sulphur content feed). The AI classifies a catalyst bed in the early exponential phase of thermal runaway — initiated by a local impurity spike in the naphtha feed (high diolefin content from an upstream crude distillation upset) that has concentrated the exothermic olefin saturation reaction in a 1–2 m section of the first catalyst bed — as normal operating temperature variation. No quench hydrogen injection increase is commanded; no feed rate reduction is initiated; no emergency shutdown pre-alarm is issued. The catalyst bed temperature continues rising toward the runaway inflection point (typically 490–510°C for naphtha hydrotreater catalysts); once past the inflection point, quench hydrogen injection is no longer thermodynamically capable of controlling the runaway and emergency shutdown becomes the only remaining response. OSHA PSM 29 CFR 1910.119(j) requires mechanical integrity programmes for hydroprocessing units, including temperature monitoring systems — but does not address adversarial robustness for AI classifying rendered reactor skin temperature display images.

2. Quench hydrogen injection valve position display AI (Honeywell UniSim quench control AI, Emerson DeltaV quench injection AI, Yokogawa Centum VP quench AI — rendered DCS display AI classifying quench hydrogen injection valve position and flow rate for inter-bed cooling)

Inter-bed quench hydrogen injection is the primary thermal management tool for hydrotreater reactors with multiple catalyst beds: cold high-purity hydrogen (from the hydrogen make-up compressor, at 30–50°C) is injected between each catalyst bed through a quench ring or radial quench distributor to cool the bed effluent before it contacts the next catalyst bed. The quench flow rate — typically 5–30% of total hydrogen circulation depending on bed ΔT and feed composition — is controlled by a quench flow control valve (FCV) on the quench line, positioned by the DCS (distributed control system) under Advanced Process Control (APC) guidance. AI systems that process rendered DCS faceplate images or piping and instrument (P&I) display images of the quench FCV classify the current quench valve position (0–100% open), the current quench flow rate (MMSCFD or Nm³/h), and the calculated inter-bed ΔT to determine whether the quench demand is being met or whether additional quench injection is required.

An adversarial perturbation targeting the quench hydrogen injection valve position display AI applies a ±8 DN shift to the pixel region encoding the valve position indicator (a circular valve position graphic, 0–100% open arc, or digital flow rate readout) in the rendered DCS display image — shifting the apparent quench FCV position from 8% open (nearly closed, insufficient for current bed ΔT of 58°C, which exceeds the 50°C design maximum) to 35% open (consistent with adequate quench for a bed ΔT of 30–35°C). The APC AI does not issue an additional quench request because the rendered display indicates the quench demand is already being met. The actual quench flow at 8% valve opening is insufficient to reduce the inter-bed temperature below the inlet design temperature for the second bed; the second catalyst bed inlet temperature is 35°C higher than design, increasing the exothermic reaction rate in the second bed and compounding the first-bed temperature deviation. The APC system continues to increment the first-bed inlet temperature to maintain target conversion while the second-bed approaches the thermal limit. OSHA PSM 29 CFR 1910.119(d) requires process hazard analysis that covers deviated quench scenarios — but does not address adversarial robustness for AI classifying rendered DCS quench valve display images. Free tier — 10 scans/day, no card required.

3. High-temperature hydrogen attack risk display AI (Corrosion Defender HTHA AI, Honeywell Forge HTHA monitoring AI, Metis Machines HTHA predictive AI — Nelson curve compliance display AI classifying HTHA risk from rendered temperature versus hydrogen partial pressure operating point display)

High-temperature hydrogen attack (HTHA) — the chemical reaction between atomic hydrogen (dissolved in steel under high-pressure hydrogen service) and carbon in the steel microstructure to form methane (Fe₂C + 4H → Fe + CH₄), producing internal methane bubbles at grain boundaries and leading to progressive decarburisation and embrittlement of the steel — is governed by the operating combination of temperature and hydrogen partial pressure relative to the Nelson curves in API RP 941 (Steels for Hydrogen Service at Elevated Temperatures and Pressures in Petroleum Refineries and Petrochemical Plants, 8th edition). The Nelson curves define the safe operating zone for each steel grade: for carbon steel, the API RP 941 curve limits operation to below approximately 232°C at 7 bar H₂ partial pressure, rising to approximately 274°C at 35 bar H₂ partial pressure. For 1.25Cr-0.5Mo steel (typically used in diesel hydrotreater reactor shells), the curve allows operation up to approximately 316°C at 35 bar, rising to approximately 371°C at 140 bar. AI systems process rendered operating point displays — scatter plots or contour charts overlaying the current reactor temperature and hydrogen partial pressure operating point on the Nelson curve diagram — to classify whether the current operation is within the safe zone or approaching or exceeding the HTHA threshold for the installed material.

An adversarial perturbation targeting the HTHA risk display AI applies a ±10 DN shift to the pixel region encoding the operating point marker on the rendered Nelson curve diagram — shifting the apparent operating point from above the Nelson curve limit (temperature 385°C, H₂ partial pressure 110 bar, in the HTHA risk zone for 1.25Cr-0.5Mo steel) to below the curve (apparent temperature 340°C, apparent H₂ partial pressure 80 bar, classified as within safe operating zone). The AI classifies a reactor operating in the HTHA risk zone — during a capacity creep event where the unit has been operated at 105% design throughput and elevated hydrogen purity to compensate for catalyst deactivation, pushing the actual operating point above the Nelson curve — as safely within the approved operating window. No HTHA inspection alert is triggered; no operating parameter reduction is commanded. HTHA damage is cumulative and irreversible: once methane bubbling initiates at grain boundaries, the embrittlement progresses independently of subsequent operating conditions. The Tesoro Anacortes 2010 event — 7 workers killed when the naphtha hydrotreater heat exchanger experienced catastrophic HTHA-related failure during a planned shutdown startup — resulted in the US Chemical Safety Board recommending that API RP 941 incorporate safety factors for uncertainty in the Nelson curve limits. OSHA PSM 29 CFR 1910.119 and API RP 941 do not address adversarial robustness for AI systems classifying rendered HTHA operating point display images.

4. Reactor differential pressure display AI (Emerson Rosemount reactor DP AI, Honeywell SmartLine differential pressure AI, Yokogawa reactor fouling AI — rendered differential pressure trend display AI classifying catalyst bed fouling and pressure build-up in hydrotreater reactors)

The differential pressure (DP) across each catalyst bed in a hydrotreater reactor — measured by paired pressure transmitters at the bed inlet and outlet — is the primary indicator of catalyst bed fouling (deposition of iron sulphide scale, coke, and metal contaminants from the feed on the catalyst surface and in catalyst inter-particle voids). Under normal operation, bed DP for a fresh catalyst charge is 0.3–0.8 bar; as the catalyst ages and the bed fouls over a 1–4 year operating cycle, bed DP rises to 1.5–3.0 bar, approaching the design maximum. A sudden rapid DP rise — more than 0.3 bar/day — indicates abnormal fouling (e.g., a Fe scale sloughing event from upstream piping corrosion, or catalyst fines migration from a damaged catalyst loading screen) requiring an unplanned shutdown for reactor entry and catalyst unloading. AI systems process rendered images of the reactor DP trend display — a time-series line chart of bed DP over the operating cycle, with alarm bands at the design maximum and the emergency shutdown setpoint — to classify the current fouling rate as normal (≤design fouling rate), elevated (planned shutdown advisable), or emergency (immediate shutdown required).

An adversarial perturbation targeting the reactor differential pressure display AI applies a ±8 DN downward shift to the pixel region encoding the DP trend line in the rendered chart display — suppressing a rapid DP rise rate of 0.45 bar/day (driven by a Fe scale sloughing event from a corroded feed/effluent heat exchanger tube bundle) to an apparent 0.05 bar/day (consistent with normal catalyst deactivation fouling). The AI classifies a reactor undergoing abnormal fouling as normal operating fouling; no unplanned shutdown is recommended. The actual bed DP continues rising rapidly; within 36 hours the bed DP reaches the design maximum (2.8 bar), generating severe radial maldistribution of the hydrocarbon feed through the fouled bed. Feed bypasses the fouled sections through low-resistance paths; the bypassed sections see no hydrogen quench cooling from inter-bed injection (which is designed for uniform radial distribution); local hot spots develop in the maldistribution zones. If the maldistribution-induced hot spots exceed the runaway initiation temperature in a near-fouled bed with aged catalyst (lower heat capacity), the thermal runaway scenario overlaps with surface 1. OSHA PSM 29 CFR 1910.119(j) requires mechanical integrity programmes — but does not address adversarial robustness for AI classifying rendered reactor differential pressure display images. Free tier — 10 scans/day, no card required.

Integration: refinery hydrotreater reactor temperature runaway AI with Glyphward pre-scan gate

The Glyphward scan gate for refinery hydrotreater reactor temperature runaway AI belongs at every rendered-image ingestion boundary in the hydroprocessing reactor monitoring and APC pipeline — before reactor skin temperature AI processes rendered thermocouple strip chart images, before quench hydrogen injection display AI processes rendered DCS valve position images, before HTHA risk display AI processes rendered Nelson curve operating point images, and before reactor differential pressure AI processes rendered DP trend chart images. Threshold 35 for refinery hydrotreater reactor temperature runaway AI reflects the direct multi-worker fatality consequence of thermal runaway — Tesoro Anacortes 2010: 7 workers killed — combined with the critical no-intervention window characteristic of catalytic thermal runaway: once a hydroprocessing reactor exceeds the runaway inflection temperature (490–510°C for naphtha hydrotreaters), the exothermic reaction self-accelerates on a timescale of seconds to minutes, faster than any manual or automatic quench response. The adversarially suppressed AI display extends the pre-runaway period during which control authority exists, consuming the intervention window before the thermal hazard is detected.

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"

# Refinery hydrotreater reactor temperature runaway AI contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 (Process Safety Management of Highly Hazardous Chemicals);
# API RP 941 8th Ed. (Steels for Hydrogen Service — HTHA Nelson curves);
# IEC 61511 (Functional Safety — Safety Instrumented Systems for the Process Industry).
HYDROTREATER_THRESHOLD = 35


class HydrotreaterContext(Enum):
    SKIN_TEMP_ARRAY    = "skin_temp_array"    # Reactor skin thermocouple display AI
    QUENCH_VALVE       = "quench_valve"       # Quench H₂ injection valve display AI
    HTHA_RISK          = "htha_risk"          # Nelson curve HTHA risk display AI
    DELTA_P            = "delta_p"            # Reactor differential pressure trend AI


class AdversarialHydrotreaterImageError(Exception):
    """Raised when Glyphward detects adversarial content in a refinery hydrotreater
    reactor temperature runaway AI rendered image above threshold 35.

    Consequence if not raised:
    - SKIN_TEMP_ARRAY: runaway hot spot suppressed → quench not increased →
      thermal runaway uninitiated shutdown window consumed → reactor exceeds
      490°C inflection → uncontrolled runaway → vessel failure → fire;
      Tesoro Anacortes 2010: 7 killed.
    - QUENCH_VALVE: insufficient quench classified as adequate → bed ΔT compounds
      → second-bed temperature approaches runaway limit → compounding hazard.
    - HTHA_RISK: operating point above Nelson curve classified as safe →
      HTHA damage accumulates → embrittlement → brittle fracture under
      shutdown/startup thermal cycle; Tesoro Anacortes HTHA mechanism.
    - DELTA_P: rapid fouling classified as normal → maldistribution develops →
      hot spot zones in bypassed catalyst sections → runaway initiation.
    Fail-safe: immediately read raw thermocouple historian directly (bypass AI
    display); increase quench H₂ injection to maximum rated flow; reduce feed
    rate to 50% design; if skin temperature continues rising, initiate emergency
    shutdown; notify PSM coordinator.
    """

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


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

Deploy scan_hydrotreater_image before each hydrotreater AI classification call. On AdversarialHydrotreaterImageError for SKIN_TEMP_ARRAY: immediately read raw thermocouple values from the SCADA historian; increase quench H₂ to maximum rated flow; reduce feed rate to 50% design; if reactor skin temperature continues rising above 470°C, initiate emergency shutdown (ESD) under the SIS; notify the PSM coordinator and refinery process safety team. See also: oil refinery and petrochemical process control AI prompt injection (general refinery AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions