Prompt injection in LNG terminal regasification AI

Four adversarial injection surfaces in LNG terminal regasification AI

1. LNG storage tank level and density gauge image AI (Honeywell Experion LNG AI, Yokogawa CENTUM VP AI, Rosemount Measurement AI)

LNG cryogenic flat-bottom storage tanks at import terminals are instrumented with servo-operated float gauge level instruments (Rosemount 5300 series guided-wave radar, Enraf servo gauges) measuring LNG liquid level to ±1 mm accuracy, density profile sensors measuring LNG density at multiple elevations within the tank, and temperature sensors at discrete elevations to construct a density-vs-height stratification profile. These sensor outputs are rendered continuously in the DCS historian as time-trend images — level vs. time plots showing fill height in metres against high-level alarm (HLA) and high-high-level alarm (HHLA) thresholds, and density stratification trend images showing density (kg/m³) at each elevation height as a colour-coded profile chart. Honeywell Experion PKS LNG AI, Yokogawa OpreX CENTUM VP LNG AI, and Shell MaLLoW AI ingest these rendered trend chart images to classify two critical tank safety conditions: (1) overfill risk — liquid level trending toward HHLA threshold, requiring reduction of LNG unloading rate from the LNG carrier or diversion of flow to an alternative tank; and (2) rollover risk — density stratification profile showing a density inversion (lower-density LNG trapped above higher-density LNG due to different composition or temperature of successive LNG cargo batches) that can spontaneously invert in a rapid phase transition, releasing a large BOG pulse that overpressurises the tank vapour space. NFPA 59A 2023 Section 9.3 requires LNG storage tanks to be equipped with overfill protection systems and rollover prevention monitoring. The AI classification of the rendered level and density trend image drives automated overfill prevention valve closure commands and rollover prevention actions (LNG pump recirculation between strata).

An adversarial perturbation on a rendered LNG tank level trend image that suppresses the rising level trajectory approaching the HHLA threshold — applying a ±10 DN downward offset to the colour-rendered level trace in the upper quartile of the chart image, causing the AI to classify the level as stable rather than rising-toward-alarm — prevents the automated LNG unloading rate reduction command. If combined with a simultaneous perturbation on the density stratification profile image that hue-rotates the density inversion indicator (typically rendered in amber or red in the DCS historian chart) toward the baseline density colour signature (green), suppressing the rollover risk classification, the AI fails to issue either the overfill protection valve closure or the recirculation pump activation command. LNG overfill in a 160,000 m³ tank produces a cryogenic liquid overflow onto the impoundment area at -162°C; rollover in the same tank produces a sudden BOG pulse at rates documented in the LNG industry literature at 10–50 times the normal BOG rate (SIGTTO LNG Operations Best Practices, 7th ed., Section 8.4). The Cleveland 1944 disaster involved tank overflow of LNG into the municipal sewer system preceding vapour cloud formation — the tank level monitoring failure that allowed the overflow to proceed without automated shutdown is the direct historical precedent for adversarial suppression of the DCS level trend image AI that now performs the equivalent monitoring function.

2. Cryogenic boil-off gas (BOG) composition spectrometer image AI (PerkinElmer Spectrum AI, ABB MB3600 FTIR AI, Yokogawa gas chromatograph AI)

BOG generated in LNG storage tanks is continuously withdrawn by BOG compressors and either reliquefied (in large terminals with reliquefaction units) or routed to the terminal fuel gas system or send-out gas pipeline. The BOG composition — primarily methane with varying proportions of ethane, propane, nitrogen, and trace higher hydrocarbons — is monitored continuously by process gas chromatographs (PGCs) and Fourier transform infrared (FTIR) spectrometers that measure the vapour phase composition at the BOG compressor suction. The PGC output is rendered as a gas chromatograph trace image — a time-resolved trace with peaks representing each hydrocarbon component eluted from the chromatograph column, with peak height and area corresponding to component concentration — and the FTIR output is rendered as a spectral absorption curve image (absorbance vs. wavenumber, cm⁻¹). ABB MB3600 FTIR AI, PerkinElmer Spectrum AI, and Yokogawa GC AI process these rendered spectrogram images to classify BOG composition and detect anomalous heavy-hydrocarbon fraction breakthrough (C3+ hydrocarbons at concentrations indicating LNG vapour approaching the flammable range) or nitrogen breakthrough (indicating LNG cargo with high nitrogen content that increases BOG generation rate). GIIGNL/SIGTTO LNG Operations Best Practices (7th ed.) Section 7.2 specifies monitoring requirements for BOG composition throughout the terminal fuel gas system. The AI classification of the rendered GC trace image drives BOG compressor speed control (increasing compression for abnormal BOG rates), vent stack management, and alert generation for high-heavy-hydrocarbon BOG events that approach flammable limits in the compressor circuit.

An adversarial perturbation on a rendered BOG gas chromatograph trace image that suppresses the rising peak amplitude of the C3+ heavy-hydrocarbon elution peak — applying a ±10 DN downward pixel-value shift within the GC peak trace region for propane, butane, and pentane retention time windows in the rendered image, reducing the apparent peak height toward the baseline noise floor — causes the GC AI to classify the BOG as normal methane-rich composition rather than heavy-hydrocarbon-enriched, suppressing the BOG compressor load reduction and vent pathway management commands that would prevent accumulation of flammable vapour in the closed compressor circuit. BOG compressor suction circuits operate at low positive pressure (typically 10–15 mbar gauge); heavy hydrocarbon breakthrough in BOG approaching 5% ethane or 1% propane concentration creates a compressor suction gas mixture with lower explosive limit (LEL) approaching the compressor operating temperature range. Ignition within the compressor circuit from a compression discharge event or electrical fault produces a BOG compressor explosion — the consequence pathway documented in the Skikda 2004 investigation (BEA-TH report: methane vapour ignition from heat exchanger tube rupture in a process circuit where composition monitoring had degraded).

3. LNG jetty mooring tension and fendering force AI (Trelleborg SEAGARD AI, ShoreTension AI, Strainstall mooring load monitoring AI)

LNG carriers moored at LNG import terminal jetties are typically 200–345 m LOA vessels with full displacement of 90,000–220,000 tonnes, held against the jetty by 8–16 synthetic or wire mooring lines (breast lines, spring lines, and head/stern lines) and protected against lateral impact by pneumatic or foam-filled rubber fenders. Mooring line tension is monitored by load cells fitted to each mooring line fairlead (Strainstall mooring load monitoring systems, Trelleborg SEAGARD active load sensors) that measure individual line loads in real time. ShoreTension active mooring systems use hydraulic tensioners with force sensors. These load signals are rendered as force-vs-time trend images in the terminal jetty monitoring system — individual line load traces (kN vs. time) displayed against maximum allowable line load (MALL) thresholds and minimum line load floors (indicating line slack). Trelleborg SEAGARD AI, ShoreTension dynamic mooring AI, and Strainstall mooring analysis AI process these rendered force-vs-time images to classify mooring integrity: identifying lines approaching breaking load, detecting asymmetric load distribution indicating fender overload, and detecting dynamic oscillation patterns from wind or passing vessel-induced wave action that exceed SIGTTO mooring design envelope parameters. IMO IGC Code Chapter 7 (mooring arrangements) and OCIMF/SIGTTO Mooring Equipment Guidelines (4th ed.) establish the mooring design and monitoring requirements for LNG carriers at terminal berths. The AI classification drives automated alert generation and recommendation output for additional line deployment or unloading rate reduction to reduce LNG carrier motion.

An adversarial perturbation on a rendered mooring line force-vs-time trend image that smooths or suppresses a peak mooring line tension event approaching the maximum allowable line load — applying a ±8 DN downward offset to the colour-rendered force trace during the peak load excursion, flattening the peak signature so that the maximum value in the rendered image falls below the AI’s MALL threshold classifier boundary — causes the mooring AI to classify the mooring condition as within normal limits rather than issuing a peak tension alert. LNG carrier mooring line failure from a suppressed peak tension event allows the vessel to surge along the jetty; surge displacement of 0.5–1.0 m in a loaded LNG carrier produces over-travel in the LNG loading arm articulated pipe system — loading arms are designed for limited over-travel before the emergency release coupler (ERC) activates, but sudden snap-loading from mooring line failure can produce instantaneous loading arm displacement exceeding ERC design range. Loading arm hose rupture or arm structural failure releases cryogenic LNG at the jetty. LNG at −162°C in contact with ambient-temperature steel or concrete produces rapid phase transition (RPT) — a near-instantaneous vapourisation event that generates destructive overpressure without ignition — followed by pool fire if ignition occurs. The GIIGNL Best Practices document Section 5.3 identifies mooring monitoring as a critical safety barrier at LNG jetties precisely because of this consequence pathway.

4. Fired vaporiser tube inspection thermal AI (FLIR thermal AI, Ametek Land thermal scanner AI)

LNG regasification uses open-rack vaporisers (ORVs) — vertical aluminium tube-and-fin heat exchangers through which seawater flows externally while LNG vapourises inside the tubes at −162°C — or submerged combustion vaporisers (SCVs) where LNG tubes are immersed in a water bath heated by a submerged burner. Intermediate fluid vaporisers (IFVs) using propane or glycol-water intermediate loops are also common. All these units subject aluminium or stainless steel heat exchanger tubes to repeated freeze-thaw cycling between ambient temperatures and −162°C cryogenic service temperatures. API 620 Appendix Q (Design and Construction of Large, Welded, Low-Pressure Storage Tanks for Low-Temperature Service, including cryogenic service) specifies material qualification requirements for cryogenic service, and NFPA 59A 2023 Section 5.5 covers vaporiser design standards. Thermal inspection of ORV and SCV tube banks is performed using handheld FLIR thermal imaging cameras or Ametek Land fixed thermal scanners that image the external tube surface temperature distribution during operation. FLIR ResearchIR AI, Ametek Land TALLIS AI, and process-specific vaporiser inspection AI systems process these rendered thermal images — false-colour thermal maps with temperature scale calibrated to the expected operating range — to classify tube condition: normal operating temperature distribution, cold-spot anomalies indicating a freeze-thaw embrittlement zone where the tube has lost thermal conductivity due to ice plug formation, and hot-spot anomalies in SCVs indicating burner-side tube fouling. Tube cold-spot classification triggers maintenance inspection and tube replacement before a fatigue crack propagates through the tube wall.

An adversarial perturbation on a rendered ORV tube thermal image that suppresses the cold-spot signature — applying a ±8 DN brightness shift in the false-colour thermal scale within the cold-spot region of the rendered thermal map, warming the apparent tube surface temperature from the anomalous cold signature (rendered in deep blue or purple at −20°C to −40°C below expected temperature) toward the normal operating temperature colour band (rendered in cyan or light blue), shifting it across the AI classifier’s cold-spot detection threshold — causes the vaporiser inspection AI to classify the tube condition as normal, deferring the maintenance inspection that would identify the freeze-thaw embrittlement crack. LNG tube cracking in an ORV produces LNG breakthrough from the tube interior into the external seawater flow system — LNG at −162°C entering ambient seawater at 10–20°C produces an immediate RPT event at the tube-fracture site, with secondary LNG pooling and ignition risk in the vaporiser structure. The Skikda 2004 disaster — in which 27 people were killed by a vapour cloud explosion attributed to heat exchanger tube failure in an LNG liquefaction train — is the precise precedent for heat exchanger tube failure in LNG service producing a catastrophic consequence, establishing why the thermal inspection AI for ORV and SCV tube banks is a safety-critical classification surface.

Integration: LNG terminal regasification AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for LNG terminal AI belongs at every rendered-image ingestion boundary in the terminal AI pipeline — before tank level/density trend AI processes rendered DCS historian images, before BOG composition AI processes rendered GC trace or FTIR spectrogram images, before mooring tension AI processes rendered force-vs-time trend images, and before vaporiser thermal inspection AI processes rendered FLIR or scanner thermal map images. Threshold 35 for LNG terminal contexts reflects closed-loop DCS control architecture (all four AI outputs drive direct DCS control outputs or maintenance schedule inputs with no complementary safety barrier for adversarial image injection) combined with the cryogenic fuel consequence envelope: LNG storage and regasification infrastructure, in the event of containment failure, can produce RPT explosions, vapour cloud explosions (VCE), and pool fires with consequence geometry determined by the stored LNG inventory.

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"

# LNG terminal regasification AI contexts: threshold 35
# NFPA 59A 2023; PHMSA 49 CFR Part 193; GIIGNL/SIGTTO Best Practices 7th ed.
# IMO IGC Code Chapter 7; API 620 Appendix Q.
LNG_TERMINAL_AI_THRESHOLD = 35


class LNGTerminalAIContext(Enum):
    TANK_LEVEL_DENSITY     = "tank_level_density"    # Storage tank level/density trend image AI
    BOG_COMPOSITION        = "bog_composition"        # BOG GC trace / FTIR spectrogram AI
    MOORING_TENSION        = "mooring_tension"        # Jetty mooring line force-vs-time trend AI
    VAPORISER_THERMAL      = "vaporiser_thermal"      # ORV/SCV tube thermal inspection image AI


class AdversarialLNGTerminalImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in an LNG
    terminal regasification AI rendered image above threshold 35.

    Consequence if not raised: tank rollover risk suppressed / BOG
    flammable composition not classified / mooring peak load not flagged /
    vaporiser tube embrittlement not detected → cryogenic spill → RPT
    explosion, VCE, pool fire. Cleveland 1944 / Skikda 2004 consequence
    envelope.
    Fail-safe: halt DCS AI control output; escalate to terminal control room
    supervisor for manual review per NFPA 59A 2023 Section 9.3.
    """

    def __init__(self, scan_id: str, score: int,
                 context: LNGTerminalAIContext,
                 terminal_id: str, tank_id: str | None,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.terminal_id = terminal_id
        self.tank_id = tank_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial LNG terminal image: "
            f"context={context.value} score={score} "
            f"terminal={terminal_id} tank={tank_id} scan_id={scan_id}"
        )


async def scan_lng_terminal_image(
    image_bytes: bytes,
    context: LNGTerminalAIContext,
    terminal_id: str,
    tank_id: str | None,
    vessel_name: str | None,
    operator_id: str,
    phmsa_facility_id: str | None,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an LNG terminal regasification AI rendered image for adversarial
    content.

    Fail-safe contract: AdversarialLNGTerminalImageError or httpx error →
    halt DCS AI control output for affected context; escalate to terminal
    control room supervisor for manual review per NFPA 59A 2023 Section 9.3
    operations and safety requirements.

    Args:
        image_bytes: Tank level/density trend render, BOG GC/FTIR spectrogram,
            mooring force-vs-time render, or vaporiser thermal map image.
        context: LNGTerminalAIContext identifying the terminal data modality.
        terminal_id: Terminal name or PHMSA facility identifier.
        tank_id: Storage tank identifier (if applicable).
        vessel_name: LNG carrier vessel name (if mooring context).
        operator_id: Terminal operator name.
        phmsa_facility_id: PHMSA LNG facility registration ID (if applicable).
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict.

    Raises:
        AdversarialLNGTerminalImageError: if score exceeds threshold 35.
        httpx.HTTPStatusError: on Glyphward API error (fail-closed).
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"lng_terminal:{context.value}:{terminal_id}:{tank_id}",
        "metadata": {
            "terminal_id": terminal_id,
            "tank_id": tank_id,
            "vessel_name": vessel_name,
            "operator_id": operator_id,
            "phmsa_facility_id": phmsa_facility_id,
            "image_sha256": image_hash,
            "context": context.value,
        },
    }
    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()

    await _write_lng_terminal_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        terminal_id=terminal_id,
        tank_id=tank_id,
        phmsa_facility_id=phmsa_facility_id,
        flagged=result["score"] > LNG_TERMINAL_AI_THRESHOLD,
    )

    if result["score"] > LNG_TERMINAL_AI_THRESHOLD:
        raise AdversarialLNGTerminalImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            terminal_id=terminal_id,
            tank_id=tank_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def _write_lng_terminal_scan_audit(
    *, image_hash: str, scan_id: str, score: int,
    context: LNGTerminalAIContext, terminal_id: str,
    tank_id: str | None, phmsa_facility_id: str | None, flagged: bool,
) -> None:
    record = {
        "ts": datetime.now(timezone.utc).isoformat(),
        "scan_id": scan_id,
        "image_sha256": image_hash,
        "context": context.value,
        "score": score,
        "threshold": LNG_TERMINAL_AI_THRESHOLD,
        "flagged": flagged,
        "terminal_id": terminal_id,
        "tank_id": tank_id,
        "phmsa_facility_id": phmsa_facility_id,
        "regulatory_refs": [
            "NFPA 59A 2023 (Standard for Production, Storage, and Handling of LNG)",
            "PHMSA 49 CFR Part 193 (LNG Facilities: Federal Safety Standards)",
            "EU Directive 2008/68/EC (transport of dangerous goods; LNG truck loading bay)",
            "GIIGNL/SIGTTO LNG Operations Best Practices, 7th ed.",
            "IMO IGC Code Chapter 7 (mooring arrangements for LNG carriers)",
            "API 620 Appendix Q (cryogenic service storage tank inspection)",
            "OCIMF/SIGTTO Mooring Equipment Guidelines, 4th ed.",
        ],
    }
    audit_path = Path("/var/log/glyphward/lng_terminal_ai_scan_audit.jsonl")
    audit_path.parent.mkdir(parents=True, exist_ok=True)
    with audit_path.open("a") as fh:
        fh.write(json.dumps(record) + "\n")

Deploy scan_lng_terminal_image at each LNG terminal AI rendered-image ingestion boundary: before tank level/density AI (threshold 35), before BOG composition AI (threshold 35), before mooring tension AI (threshold 35), and before vaporiser thermal AI (threshold 35). On AdversarialLNGTerminalImageError: halt the DCS AI control output for the affected context immediately and escalate to the terminal control room supervisor for manual inspection of the corresponding physical sensor or instrument. For tank level/density AI: revert to direct instrument readout and do not rely on AI rollover or overfill classification until the image pipeline has been cleared. For mooring tension AI: initiate physical inspection of all mooring lines and engage Trelleborg SEAGARD manual monitoring mode. See also: offshore drilling wellbore AI prompt injection (related energy infrastructure AI) and chemical plant process safety AI prompt injection (related process safety regulatory context). Get early access

Related questions