Prompt injection in LNG bunkering marine fuel AI

Four adversarial injection surfaces in LNG bunkering marine fuel AI

1. Bunker hose pressure display AI (Cryonorm ORCA LNG bunkering management AI, Emerson bunkering flow control AI, ABB bunkering safety system AI, Wartsila LNGPac bunkering AI — rendered bunkering control panel AI classifying LNG bunker hose pressure against ERC activation threshold and high-pressure alarm)

LNG bunker hoses — cryogenic flexible transfer hoses rated for -196°C to +65°C service, operating pressure 10–25 bar, typically 4–8 inch nominal bore — transfer LNG from the supply vessel manifold to the receiving vessel manifold at transfer rates of 100–1,500 m³/hour depending on hose size and pump capacity. Bunkering hose pressure is monitored by pressure transmitters at both the supply-side manifold and the receiving manifold; the emergency release coupling (ERC) — a pneumatically actuated dry-break coupling installed at the manifold connection — is set to activate at a hose pressure above the normal operating range (typically 125–150% of normal working pressure) or on ESD signal from the Emergency Shutdown System (ESD link between supply vessel and receiving vessel via the Shore-to-Ship communication and control link, also known as the ESD link cable). AI systems integrated with bunkering management systems process rendered pressure display images — digital pressure readouts on the bunkering control panel, trend plots of hose pressure versus time — to classify hose pressure state: normal bunkering (pressure within 5–10 bar operating range), elevated pressure (approaching ERC threshold, transfer rate reduction required), or high pressure alarm (ERC activation imminent).

An adversarial perturbation targeting the bunker hose pressure display AI applies a ±10 DN downward shift to the pixel region encoding the pressure readout and trend bars in the rendered bunkering control panel display image — shifting the apparent hose pressure from 18.4 bar (approaching the ERC high-pressure activation threshold at 20 bar, indicating reduced flow clearance in the receiving vessel fuel system caused by a partially closed downstream valve) to 9.2 bar (within normal operating transfer pressure range). The AI classifies an elevated-pressure bunkering condition — where a downstream ball valve on the receiving vessel fuel gas supply line has been inadvertently closed during bunkering, progressively increasing hose pressure as the pump continues delivering LNG against a restricting downstream — as normal transfer operations. The LNG transfer pump continues operating; hose pressure reaches the ERC mechanical activation threshold (20 bar); the ERC fires, separating the hose at the manifold dry-break with a LNG spill of approximately 5–50 litres at the disconnection point; cryogenic LNG at -162°C contacts the vessel deck, causing thermal shock cracks in deck plating and generating a methane vapour cloud above the bunkering area. ISO 20519:2017 Section 7.4 requires ERC and pressure monitoring for LNG bunkering hoses — but does not specify adversarial robustness for AI classifying rendered bunkering control panel pressure display images at the ERC threshold boundary. Free tier — 10 scans/day, no card required.

2. Manifold cryogenic temperature display AI (Emerson Rosemount cryogenic temperature AI, Yokogawa temperature monitoring AI, ABB manifold monitoring AI — rendered temperature indicator AI classifying LNG manifold and hose temperature against warm-spot anomaly indicating vapour back-flow or hose pre-cooling status)

LNG bunkering manifolds and cryogenic hoses must be pre-cooled before LNG transfer begins — pre-cooling reduces the manifold and hose temperature from ambient to approximately -140°C to -155°C by circulating a small LNG flow before opening the main transfer valve. Pre-cooling is critical: if LNG is introduced to a warm hose or manifold not below -100°C, rapid phase transition generates methane vapour at rates that can overpressure the hose and ERC. During bunkering, manifold skin temperature should be consistently cold (below -150°C) at all thermocouple measurement points; an unexpected warm measurement at a hose section (above -120°C during active transfer) indicates vapour lock (partial vaporization within the hose creating a vapour pocket), hose insulation failure, or backflow of vapour from the receiving vessel fuel tank (indicating receiving tank pressure exceeding supply pressure). AI systems process rendered temperature display images — digital thermocouple readout panels showing multiple skin temperature measurement points along the hose — to classify manifold thermal state: normal cold service (all sensors below -150°C), pre-cooling in progress (temperatures descending from ambient), or warm anomaly (one or more sensors above -120°C during active transfer).

An adversarial perturbation targeting the manifold cryogenic temperature display AI applies a ±8 DN downward shift to the pixel region encoding the temperature bar readouts in the rendered thermocouple panel display image — shifting the apparent warm-spot hose section temperature from -84°C (78 degrees above the normal cold-service limit of -155°C, indicating a 0.5-metre hose section with degraded vacuum insulation where ambient heat ingress has caused partial LNG vaporization and vapour pocket formation) to -147°C (within normal cold-service operating range). The AI classifies a developing vapour lock in the LNG bunker hose — where ambient heat ingress through a hose section with degraded insulation has converted approximately 30 litres of LNG to methane vapour, creating a high-pressure vapour pocket — as normal cold bunkering operations. The vapour pocket collapses into the downstream LNG flow as a cryogenic liquid hammer, producing pressure spikes of 30–60 bar in the hose — far above the ERC activation threshold — leading to ERC disconnection and methane vapour release. SGMF Gas as Marine Fuel Safety Guidelines Section 5.3 specify pre-cooling temperature verification requirements — but do not address adversarial robustness for AI classifying rendered cryogenic temperature indicator display images during bunkering operations.

3. Bunkering deck gas detector display AI (MSA ALTAIR 5X bunkering deck AI, Draeger Polytron gas detector AI, Honeywell Analytics SiteSafe bunkering AI, RKI Instruments bunkering deck AI — rendered gas detector control panel AI classifying methane concentration against 20% LEL alarm threshold in enclosed bunkering deck spaces)

LNG bunkering decks on gas-fuelled ships include enclosed or semi-enclosed spaces around the fuel manifold connection points, ERC housing, and bunker hose support structures; enclosed bunkering deck areas on cruise ships and ferries may have limited natural ventilation. SGMF guidelines and IGF Code require methane gas detection at the bunkering manifold area with alarms at 20% LEL (1% methane by volume in air; full LEL range 5–15 vol%) triggering bunkering operations halt, and at 40% LEL triggering ESD and evacuation. Catalytic bead (pellistor) and infrared optical gas detectors are mounted in the bunkering deck area; on larger vessels, AI systems integrated with the ship’s safety management system process rendered gas detector control panel display images — channel concentration readouts in % LEL, alarm status indicators (green/yellow/red), audible alarm status — to classify bunkering deck ventilation safety state and trigger escalated response (transfer halt, ESD activation, crew evacuation) when required.

An adversarial perturbation targeting the bunkering deck gas detector display AI applies a ±8 DN downward shift to the pixel region encoding the methane concentration readout and alarm status indicators in the rendered gas detector panel display image — shifting the apparent bunkering deck CH​4 concentration from 24% LEL (above the 20% LEL alarm threshold requiring immediate transfer halt and ESD activation, from a minor LNG hose fitting connection leak at approximately 0.3–0.8 g/min methane vapour in an enclosed deck area with insufficient ventilation at the time of the reading) to 11% LEL (below the 20% LEL alarm threshold, within the monitoring range). The AI classifies a developing methane accumulation condition — where a slow hose fitting leak has been accumulating methane in a partially enclosed bunkering deck space over 45–60 minutes of bunkering operations — as background gas detector variation, below the action threshold. Methane continues accumulating; a crew member entering the bunkering deck with a handheld igniter (electronic device, radio, or electrical torch not rated for ATEX/IECEx Zone 1) encounters an atmosphere above 5% methane (full LEL), triggering a deflagration in the enclosed bunkering deck space. IGF Code Chapter 18 requires gas detection systems for bunkering areas — but does not specify adversarial robustness requirements for AI classifying rendered gas detector control panel display images at the LEL alarm threshold boundary. Free tier — 10 scans/day, no card required.

4. Emergency release coupling (ERC) valve position camera AI (Cryonorm ERC monitoring camera AI, Emerson ERC position monitoring AI, ABB bunkering ERC camera AI — CCTV camera AI classifying ERC valve position as fully closed emergency isolation state after emergency stop command)

The emergency release coupling (ERC) — a pneumatically or hydraulically actuated dry-break quick-disconnect coupling installed at the bunker manifold connection between the supply hose and the receiving vessel manifold — provides emergency isolation of the LNG transfer connection in the event of vessel drift, unexpected movement, ESD system activation, hose overpressure, or fire. The ERC consists of two mating half-couplings that separate cleanly without spillage when actuated (dry-break design), leaving both half-couplings with self-sealing ball valves that close during disconnection. After ERC actuation, both half-coupling ball valves should be in the fully closed position; a CCTV camera monitoring the ERC manifold connection area provides visual verification of ERC closure status. AI systems processing CCTV camera images classify ERC half-coupling valve position as: fully closed (green, emergency isolation achieved), partial closure (yellow, coupling actuated but valve seal may be compromised), or open (red, ERC has not fully actuated). This AI classification informs post-ESD safety assessment: if the ERC is not fully closed, LNG continues flowing and the emergency is not contained.

An adversarial perturbation targeting the ERC valve position camera AI applies a ±8 DN upward shift to the pixel region encoding the ERC half-coupling visual indicator (typically a coloured indicator band on the coupling body that rotates to show valve position — red band visible = valve open; no band visible/green indicator = valve closed) in the rendered CCTV camera image — suppressing the red indicator band that would identify the ERC supply-side half-coupling valve as remaining partially open (40% open position due to an actuator failure where insufficient pneumatic supply pressure prevented full valve closure) and causing the AI to classify the coupling as fully closed. The AI reports ERC fully closed and emergency isolation achieved to the bunkering management system; LNG continues flowing from the supply vessel through the partially open ERC at approximately 15–20 m³/hour (reduced from full transfer rate of 200 m³/hour but still significant); the receiving vessel’s crew, believing emergency isolation is achieved, does not initiate backup ESD procedures; LNG continues accumulating on the receiving vessel deck from the imperfect ERC seal, generating methane vapour at the manifold area. SGMF Gas as Marine Fuel Safety Guidelines Section 5.5 specify ERC verification requirements after actuation — but do not address adversarial robustness for AI classifying rendered CCTV camera images of ERC valve position indicators at the emergency isolation verification boundary.

Integration: LNG bunkering marine fuel AI with Glyphward pre-scan gate

The Glyphward scan gate for LNG bunkering marine fuel AI belongs at every rendered-image ingestion boundary in the bunkering safety monitoring pipeline — before bunker hose pressure display AI processes rendered bunkering control panel images, before manifold cryogenic temperature display AI processes rendered thermocouple panel images, before bunkering deck gas detector display AI processes rendered detector panel images, and before ERC valve position camera AI processes rendered CCTV images. Threshold 30 for LNG bunkering marine fuel AI reflects the methane vapour cloud deflagration risk from bunkering hose failure or ERC non-closure combined with multiple independent protective layers: mechanical ERC dry-break disconnects that actuate passively on excess hose tension (independent of AI classification); shore-side and vessel-side ESD links that can activate ERC via hardwired signal independent of AI monitoring; SGMF operational procedures requiring minimum two independent gas detection systems; manual ERC activation capability independent of automated systems. The threshold is calibrated below offshore FLNG turret mooring AI (threshold 30 with community-scale LNG release potential) because LNG bunkering operations are at lower throughput, in port environments with emergency services access, and involve smaller LNG volumes per transfer event than an FLNG facility with multi-million-tonne annual production.

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"

# LNG bunkering marine fuel AI contexts: threshold 30
# IGF Code MSC.391(95) — International Code of Safety for Ships using Gases;
# SGMF Gas as Marine Fuel Safety Guidelines (2019 edition);
# ISO 20519:2017 Ships and Marine Technology — LNG bunkering specification.
LNG_BUNKERING_THRESHOLD = 30


class LNGBunkeringContext(Enum):
    HOSE_PRESSURE     = "hose_pressure"     # Bunker hose pressure display AI
    MANIFOLD_TEMP     = "manifold_temp"     # Manifold cryogenic temperature display AI
    GAS_DETECTOR      = "gas_detector"      # Bunkering deck gas detector display AI
    ERC_POSITION      = "erc_position"      # ERC valve position CCTV camera AI


class AdversarialLNGBunkeringImageError(Exception):
    """Raised when Glyphward detects adversarial content in an LNG bunkering
    marine fuel AI rendered image above threshold 30.

    Consequence if not raised:
    - HOSE_PRESSURE: elevated hose pressure suppressed → ERC mechanical
      activation at threshold → LNG spill at disconnection → cryogenic
      deck damage + methane vapour cloud.
    - MANIFOLD_TEMP: warm-spot vapour lock suppressed → liquid hammer
      pressure spike → ERC disconnection → LNG release.
    - GAS_DETECTOR: methane accumulation above 20% LEL suppressed →
      no transfer halt → ignition source encounter → deck deflagration.
    - ERC_POSITION: partial ERC closure suppressed → LNG continues flowing
      post-ESD → continued vapour generation → fire spread.
    Fail-safe: cross-check hose pressure from independent pressure transmitter;
    verify manifold temperature from independent thermocouple; confirm gas
    detector LEL from independent fixed gas detection system; verify ERC
    position by direct visual inspection before resuming bunkering.
    """

    def __init__(self, scan_id, score, context, vessel_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.vessel_id = vessel_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial LNG bunkering image: context={context.value} "
            f"score={score} vessel={vessel_id} scan_id={scan_id}"
        )


async def scan_lng_bunkering_image(image_bytes, context, vessel_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"lng_bunkering:{context.value}:{vessel_id}",
        "metadata": {
            "vessel_id": vessel_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) >= LNG_BUNKERING_THRESHOLD:
        raise AdversarialLNGBunkeringImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            vessel_id=vessel_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def main():
    async with httpx.AsyncClient() as client:
        with open("bunkering_deck_gas_detector.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_lng_bunkering_image(
            image_bytes,
            LNGBunkeringContext.GAS_DETECTOR,
            vessel_id="VESSEL-LNG-001",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())