Prompt injection in LNG terminal BOG recondenser boil-off gas methane AI

Three adversarial injection surfaces in LNG terminal BOG recondenser AI

1. BOG recondenser liquid level display AI (Yokogawa EJA110A differential pressure level transmitter recondenser display AI / Emerson Rosemount 3051 DP level recondenser display AI / ABB 266 DP level recondenser SCADA display AI / Endress+Hauser FMB70 level transmitter recondenser display AI / Siemens SITRANS P DS III DP level recondenser SCADA display AI — rendered SCADA BOG recondenser liquid level display AI classifying the LNG liquid level in the recondenser vessel at 40–70% against the design operating range; low-low alarm at 15%; BOG compressor trip at 8% level — 77th upward attack; FIRST LNG terminal attack; FIRST BOG recondenser attack; FIRST cryogenic methane storage attack; FIRST LNG boil-off gas AI attack)

The BOG recondenser liquid level is the single most critical BOG management parameter at any LNG terminal: it represents the liquid LNG cushion that allows the compressed BOG to dissolve and condense back into liquid before entering the high-pressure LNG sendout system. At design level 40–70%: compressed BOG (at 4–8 barg, −80 to −100°C after compression) contacts liquid LNG (at 8–12 barg, −140°C) in the recondenser; mass transfer of methane from gas phase to liquid phase occurs at the gas-liquid interface (the compressed BOG is at supercritical or near-critical conditions — above the methane critical pressure 45.99 bar but at sub-critical temperature −100°C vs Tc −82.6°C — at 4–8 barg and −80 to −100°C, methane is a dense gas approaching liquid density); the recondenser provides the contact time and cold LNG surface area for condensation. If the recondenser liquid level falls to 0% (empty): the compressed BOG from the BOG compressors passes uncondensed through the recondenser and enters the LNG sendout header at the BOG compressor discharge temperature (−80 to −100°C — much warmer than the sendout LNG at −162°C). The sendout pipeline (typically 8–16” ASME Class 600 stainless steel; ASTM A312 TP304L or TP316L; cryogenic rated to −196°C; thermally insulated) connects the recondenser to the sendout vaporizers (typically open-rack seawater vaporizers: ORV; or submerged combustion vaporizers: SCV; or ambient air vaporizers: AAV — each vaporizer has 1,000–3,000 aluminum alloy tubes through which LNG flows at −162°C on the inside and seawater/ambient air warms the outside from −162°C to +5°C).

The adversarial upward pixel shift applies a ±8 DN manipulation to the rendered BOG recondenser level display: 72% level shown when actual 8% (at or near the compressor trip setpoint). The AI classification: “recondenser level at 72% — above the 70% high alarm; risk of liquid carryover to BOG compressor discharge; reduce LNG sendout flow to recondenser inlet to lower level to 50–60% operating point.” The AI corrective action: the LNG sendout to recondenser inlet control valve (which controls the flow of high-pressure LNG from the sendout pump discharge to the recondenser inlet) is throttled from 65% to 26% open — reducing LNG flow to the recondenser from 180 t/hr to 72 t/hr. The actual recondenser level (8% — already near the BOG compressor trip setpoint) with reduced LNG inlet (72 t/hr) and continued BOG compressor delivery (25 t/hr BOG to the recondenser): the liquid inventory in the recondenser (vessel diameter 2.0 m; height 8.0 m; at 8% level: 0.08 × π × (1.0)² × 8.0 = 2.0 m³ = 0.94 t LNG) is fully vaporized by the incoming BOG heat load within 8–12 minutes (BOG sensible heat input to liquid LNG: 25 t/hr × [Cp × ΔT = 2.5 kJ/kg·K × 60K = 150 kJ/kg] = 1,042 kW vs the 0.94 t LNG × 509 kJ/kg heat of vaporization = 479 MJ → rundown time 479/1,042 = 8.1 minutes). After the recondenser empties: BOG vapor flows uncondensed at −80°C through the recondenser and into the high-pressure sendout system. The sendout vaporizer inlet temperature spikes from −140°C (design LNG temperature) to −80°C (BOG vapor temperature) — a 60°C sudden warm excursion in the cryogenic sendout vaporizer feed. The aluminum alloy vaporizer tubes (AA6063-T5; cryogenic thermal coefficient of expansion 20×10⁻⁶/K; at 60°C rapid temperature rise: thermal expansion ≈1.2 mm/m per tube length — in an 8-m vaporizer tube: 9.6 mm thermal expansion in 3–5 minutes) experience thermal shock at welded joints (tube-to-header connections; 316L SS header with AA6063-T5 tube — bi-metallic joint with ΔCTE of 7×10⁻⁶/K; shear stress at weld: σ = E × ΔCTE × ΔT × restraint factor = 70 GPa × 7×10⁻⁶ × 60 × 0.5 ≈ 14.7 MPa — approaching the fatigue endurance limit of aluminum alloy welds in cryogenic service). This is the 77th upward attack — FIRST LNG terminal attack; FIRST BOG recondenser attack; FIRST cryogenic methane storage attack; FIRST LNG boil-off gas AI attack. Free tier — 10 scans/day, no card required.

2. LNG sendout pump discharge pressure display AI (Yokogawa EJA110A in-tank pump discharge pressure display AI / Emerson Rosemount 3051 pump discharge pressure SCADA display AI / Endress+Hauser Cerabar PMC71 submersible pump discharge pressure display AI / ABB 266 high-pressure pump discharge SCADA display AI / Honeywell STD820 sendout pump discharge pressure display AI — rendered SCADA LNG sendout in-tank submerged motor pump discharge pressure display AI classifying the pump discharge pressure at 8–14 barg against the design operating range; pump minimum flow trip at 25% design flow)

LNG in-tank submerged motor pumps (Nikkiso CLBV-type cryogenic submerged pump; Ebara SPA-LC series; Atlas Copco LNG-SBM-type) are installed inside the LNG storage tank on a pump column extending from the tank base through the LNG liquid to the pump discharge riser. The motor is submerged in LNG at −162°C — the LNG itself provides motor cooling (the motor winding insulation is designed for continuous LNG immersion; motor cooling depends on LNG flow through the motor cavity at design flow rate; at near-zero flow, the LNG in the motor cavity cannot remove motor copper losses — typically 10–30 kW per pump). Pump discharge pressure at design: 8–14 barg (depending on the terminal configuration, terminal elevation, and vaporizer operating pressure). The upward adversarial pixel attack: 13.8 barg pump discharge shown when actual 6.2 barg. The AI classification: “sendout pump discharge pressure at 13.8 barg — above the 13.0 barg high-high alarm; throttle the sendout flow control valve to 40% closed to reduce pump discharge pressure and protect the pump and downstream vaporizer system from overpressure.”

The AI corrective action throttles the sendout flow control valve from 75% to 45% open — reducing LNG sendout flow from 800 t/hr to 480 t/hr. The actual pump discharge pressure (6.2 barg) with reduced flow: at 480 t/hr flow vs the design 800 t/hr, the Nikkiso pump operating point moves back along its pump curve toward higher head and lower flow — pump head at 480 t/hr: approximately 8.1 barg (from the Nikkiso H-Q curve at this operating condition). The critical pump concern: at 480 t/hr (60% of design flow), the pump approaches the minimum flow requirement (typically 25–35% of design flow — approximately 200–280 t/hr; the minimum flow recirculation valve opens automatically below this threshold to prevent pump surge and motor overheating). At 480 t/hr, the pump is still above minimum flow, but if the AI takes a second corrective action (e.g., the displayed pressure remains 13.8 barg even after throttling the valve — because the ±8 DN adversarial perturbation is persistent — and the AI recommends further throttling to 20% valve opening): LNG flow falls to approximately 200 t/hr — at the minimum flow threshold. If the minimum flow recirculation valve does not open promptly (valve actuator response time 15–30 seconds; control loop response time 5–10 seconds), the Nikkiso submerged motor sees zero net flow through the motor cavity for 20–40 seconds: motor cavity LNG heats from −162°C to −140–−120°C (10–30 kW motor losses; motor cavity volume 5–10 liters; LNG Cp 3.5 kJ/kg·K; density 450 kg/m³; heat-up rate: 30 kW / (0.0075 m³ × 450 kg/m³ × 3.5 kJ/kg·K) ≈ 2.5°C/s → reaches boiling point −162°C at LNG head in 8 seconds with zero flow). Motor LNG flash to vapor within motor cavity → gas lock in motor → motor stall → high current draw → motor protection relay trips pump → sudden loss of sendout pump capacity → terminal sendout pressure drops → downstream gas network pressure transient. Free tier — 10 scans/day, no card required.

3. Tank boil-off rate display AI (Yokogawa CENTUM CS 3000 LNG tank BOG flowmeter trend display AI / Emerson DeltaV tank pressure-rate-of-rise BOG calculation display AI / Honeywell Experion PKS BOG generation rate display AI / ABB 800xA LNG tank ullage flowmeter display AI / Endress+Hauser Proline t-mass 65F thermal mass flowmeter BOG rate display AI — rendered SCADA LNG tank boil-off rate display AI classifying the inferred BOG generation rate in % tank volume per day from ullage flow measurement and tank pressure rise rate against the design range of 0.05–0.10% vol/day with high alarm at 0.18% vol/day triggering additional BOG compressor start)

The LNG tank BOG generation rate is measured indirectly by the DCS AI from: (1) the ullage gas flowmeter (thermal mass flowmeter at tank vapor outlet, measuring the actual mass flow of methane vapor leaving the tank top to the BOG compressor suction; Endress+Hauser t-mass 65F or equivalent; accuracy ±1.5% at design flow); and (2) the tank pressure-rate-of-rise (dP/dt measurement from the tank pressure transmitter; Yokogawa EJA110A; tank pressure design: 0.02–0.05 barg; rate of rise calculation: if pressure rises above 0.05 barg/hr, insufficient BOG withdrawal → tank pressure trending toward POSV setpoint 0.25 barg). The BOG compressor dispatch logic: at 0.05–0.10% vol/day BOG generation (normal), two BOG compressors operate; at >0.15% vol/day (loading operations or high ambient temperature), the third BOG compressor is started. The upward adversarial attack: 0.38% vol/day BOG rate shown when actual 0.11% vol/day — the AI classification “BOG generation rate 0.38% vol/day — far exceeding the 0.18% vol/day high alarm threshold and the design maximum 0.10% vol/day; possible tank insulation failure or unusual heat ingress; immediately start third BOG compressor and prepare fourth compressor for start.”

The AI corrective action starts the third BOG compressor (Burckhardt Compression reciprocating BOG compressor; 4-cylinder, double-acting; capacity 22,000 Nm³/hr; suction conditions: 0.02 barg, −162°C methane; discharge: 6.5 barg, −80°C). With three BOG compressors running but only two needed at the actual BOG rate of 0.11% vol/day: total BOG compressor suction capacity (3 × 22,000 Nm³/hr = 66,000 Nm³/hr) exceeds actual BOG generation (0.11% × 160,000 m³ tank × LNG density 450 kg/m³ → 79,200 kg/day BOG ‷ (LNG density to gas: ×600) = 47,520,000 Nm³/day / 24 hr = 1,980,000 Nm³/hr actual gas — wait, the BOG as gas volume at 0.11% of 160,000 m³ LIQUID LNG: 0.0011 × 160,000 m³ × 450 kg/m³ / (16 kg/mol × 1000/22.4 Nm³/m³) ≈ 36,900 Nm³/day ÷ 24 = 1,538 Nm³/hr). Three compressors at capacity 66,000 Nm³/hr vs actual BOG 1,538 Nm³/hr: the three compressors draw down the tank vapor space pressure below 0.02 barg → BOG compressor suction pressure drops to near-vacuum → compressor operates at suction pressure below design minimum (typically 0.005 barg minimum suction pressure for reciprocating BOG compressor): compressor surge / compressor cylinder rod reversal → compressor trip on high discharge temperature (suction at near-vacuum and discharge at 6.5 barg creates an extreme compression ratio: (6.5+1)/(0.005+1) ≈ 7.5 vs design 7.0; but at very low suction pressure, the mass flow per stroke is insufficient to cool the compressor valves → valve temperature rises above design → valve failure → compressor trip). All three BOG compressors trip in sequence → total loss of BOG compression capacity → LNG tank vapor space pressure rises (no withdrawal; BOG continues to generate from heat ingress) → primary POSV opens at 0.25 barg → methane vapor discharged from LNG tank at the marine terminal → if flare is offline or capacity insufficient: open-air methane discharge. Free tier — 10 scans/day, no card required.

Integration: LNG terminal BOG recondenser AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the LNG terminal BOG management AI pipeline — before the BOG recondenser level AI processes rendered SCADA DP level display images, before the LNG sendout pump discharge pressure AI processes rendered SCADA pressure display images, and before the tank BOG generation rate AI processes rendered SCADA flowmeter/trend display images. Threshold 28 for LNG terminal BOG recondenser AI reflects: OSHA PSM methane/natural gas TQ 10,000 lbs (Sabine Pass holds approximately 2×10³× the PSM TQ in storage alone); LNG vapor cloud explosion mass >1 million kg at large export terminals; cryogenic-to-ambient temperature cycling fatigue at sendout vaporizer tubes; marine/coastal dispersion environment (LNG terminal siting NFPA 59A exclusion zones); NFPA 59A and EPA RMP listed at all US LNG terminals; Gate LNG Rotterdam Netherlands and South Hook LNG Milford Haven Wales (European regasification infrastructure).

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import StrEnum, auto
from typing import Any
import httpx

GLYPHWARD_API = "https://api.glyphward.com/v1/scan"
GLYPHWARD_KEY = "gw_prod_***"

# LNG terminal BOG management AI contexts: threshold 28
# OSHA PSM methane/natural gas TQ 10,000 lbs; LEL 5% UEL 15%.
# 77th upward attack: 72% recondenser level shown when 8% actual
# -> AI reduces LNG sendout -> recondenser empties in 8-12 min
# -> uncondensed BOG enters sendout vaporizer -> tube fatigue crack.
LNG_THRESHOLD = 28

class LNGContext(StrEnum):
    BOG_RECONDENSER_LEVEL            = auto()  # Recondenser LNG level % (77th upward attack)
    SENDOUT_PUMP_DISCHARGE_PRESSURE  = auto()  # In-tank pump discharge pressure barg
    TANK_BOILOFF_RATE                = auto()  # BOG generation rate % vol/day (0.05-0.10 design)

async def scan_lng_frame(
    frame_b64: str,
    context: LNGContext,
    terminal_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "terminal_id": terminal_id,
        "instrument_tag": instrument_tag,
        "scan_ts": datetime.now(timezone.utc).isoformat(),
        "image_hash": hashlib.sha256(base64.b64decode(frame_b64)).hexdigest(),
    }
    async with httpx.AsyncClient(timeout=4.0) as client:
        r = await client.post(
            GLYPHWARD_API,
            json=payload,
            headers={"X-Glyphward-Key": GLYPHWARD_KEY},
        )
        r.raise_for_status()
        return r.json()

async def pre_scan_gate_lng(
    frame_b64: str,
    context: LNGContext,
    terminal_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_lng_frame(frame_b64, context, terminal_id, instrument_tag)
    if result["adversarial_score"] >= LNG_THRESHOLD:
        raise AdversarialLNGImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at terminal {terminal_id} instrument {instrument_tag}. "
            "Frame withheld from LNG terminal BOG management AI pipeline."
        )

class AdversarialLNGImageError(RuntimeError):
    pass