Prompt injection in urban gas distribution city gate station AI

Four adversarial injection surfaces in urban gas distribution city gate station AI

1. Pressure regulator downstream display AI (Emerson Fisher FIELDVUE DVC6200 pressure AI, Fisher BPC regulator AI, Sensus SRX district pressure AI)

The city gate station district pressure regulator — the primary pressure reduction valve that reduces the transmission system delivery pressure (200–1,400 psi) to the distribution system operating pressure (0.25–2 psi low-pressure, 2–60 psi medium-pressure) — is the most safety-critical component in the urban gas distribution pressure management chain. The district regulator is typically a spring-loaded or pilot-operated pressure regulator with a loading pilot that senses the downstream distribution pressure through a signal line and modulates the main valve position to maintain downstream pressure within the design setpoint (for low-pressure systems: 0.25 psi ± 0.05 psi; for medium-pressure: 2 psi ± 0.5 psi). Residential gas infrastructure — iron or steel distribution mains (operating at 0.25–2 psi), polyethylene service lines (rated to 5 psi, SDR 11 PE), and gas meters and appliance regulators (rated to 0.5–2 psi design pressure) — is not designed to withstand pressures above approximately 5–10 psi without joint failure or appliance bypass leakage. Over-pressurisation of the residential distribution system to 60–75 psi — as occurred in the Columbia Gas 2018 Merrimack Valley disaster — causes gas to bypass the appliance regulator at every appliance in the affected service zone, flooding the interior of connected homes with natural gas (methane) at volumes sufficient to reach the explosive range (LEL 5%, UEL 15%) within minutes. Downstream pressure AI systems process rendered pressure gauge digital displays — SCADA or local display screens showing downstream distribution pressure in psi or kPa, trend history, and alarm status — to classify distribution pressure condition: within setpoint (pressure within ±20% of design setpoint, normal regulation), elevated (pressure above setpoint — regulator inspection required), high-pressure alarm (pressure above maximum design operating pressure (MAOP) of distribution system — regulator shut-in and emergency maintenance required), and over-pressurization (pressure approaching residential appliance design limit — immediate isolation valve closure and emergency response required).

An adversarial perturbation on a rendered downstream pressure regulator SCADA display that suppresses an over-pressurization indication — applying a ±10 DN downward shift in the pixel region encoding the digital downstream pressure readout or trend trace (reducing the apparent downstream pressure from the over-pressure or high-alarm range to the within-setpoint range) — causes the pressure regulator AI to classify an active over-pressurization event as normal distribution pressure regulation, suppressing the regulator inspection and isolation valve closure that a high-pressure or over-pressurization classification would require. With the downstream pressure above MAOP and increasing without automated isolation response, the residential distribution mains and service lines throughout the affected zone experience pressure-induced gas joint separation at cast iron bell-and-spigot joints (bell-and-spigot cast iron joints — the joint type used in approximately 200,000 miles of US gas distribution mains installed before 1970 — are rated for 0.25–2 psi distribution pressure and typically separate under sustained over-pressurization above 5–10 psi), and appliance regulators bypass at each connected appliance, releasing gas into the interior of every home in the affected distribution zone simultaneously. The Columbia Gas 2018 disaster (Lawrence, Andover, North Andover, Massachusetts) resulted in more than 80 house fires and 5 structure destructions when gas released from over-pressurized appliances ignited at pilot lights and appliance burners within minutes of the over-pressurization event; 8,600 customers were without gas service for the 2018–2019 heating season as Columbia Gas replaced the entire affected distribution infrastructure at a cost exceeding $143 million. NTSB PAR-19-02 identified the pressure regulator signal line connection error as the initiating cause — a cause that an automated downstream pressure AI with adversarial injection protection could have detected and escalated in the seconds before the distribution pressure rose above MAOP.

2. Gas chromatograph Wobbe index and Btu content display AI (ABB Totalflow GC AI, Emerson NGC 8200 gas chromatograph AI, Siemens SITRANS CV process gas chromatograph AI)

The Wobbe index — a measure of gas interchangeability defined as the ratio of the higher heating value (HHV, in BTU/standard cubic foot) to the square root of the specific gravity relative to air (Wobbe index = HHV / √SG) — is the primary indicator of whether a natural gas supply will combust stably and at design heat output in residential and commercial gas appliances. US natural gas distributed through LDC networks has a typical Wobbe index range of approximately 1,310–1,390 BTU/SCF (Interchangeability Class 1; ASTM D1945 composition specification), with a higher heating value of approximately 1,010–1,075 BTU/SCF and specific gravity of approximately 0.58–0.62 relative to air. Residential appliances (ranges, furnaces, water heaters, dryers) are designed for this Wobbe index range — their gas orifice sizes, burner design, and combustion air supply are fixed at manufacture to achieve the design heat output, flame stability, and combustion efficiency at Wobbe indices within approximately ±5% of the appliance’s design specification. If the Wobbe index of the distributed gas deviates outside the appliance design range — due to injection of N2 or CO2-rich gas (from CCS or biogas injection into the network, reducing Wobbe index), injection of high-LPG gas (from peaking supply, increasing Wobbe index), or contamination of the gas stream with inert dilutants — residential appliances experience off-design combustion: high Wobbe index gas (rich combustion) causes flashback (flame burning back into the appliance burner port, creating localized overheating and appliance fire risk); low Wobbe index gas (lean combustion) produces elevated CO (carbon monoxide) at sub-stoichiometric air:fuel ratios because the appliance combustion air inlet is fixed and does not adjust for the lower Wobbe gas. Gas chromatographs at city gate stations measure the gas composition (C1–C6+, N2, CO2) and compute the Wobbe index, HHV, and specific gravity in real time, rendering the output as a digital display screen showing current values, trend history, and specification alarm status. AI systems process rendered GC display images to classify gas quality: within specification (Wobbe index within AGA Interchangeability window, HHV within pipeline quality specification), approaching limit (trending toward specification boundary — supply review required), specification deviation (Wobbe index outside specification — gas diversion or blending required before distribution), and off-specification (Wobbe severely outside range — immediate distribution system isolation required).

An adversarial perturbation on a rendered gas chromatograph display that suppresses a Wobbe index deviation — applying a ±8 DN shift to the pixel region encoding the digital Wobbe index value or trend trace (normalising the apparent Wobbe reading from the off-specification range to within the AGA Interchangeability window) — causes the GC monitoring AI to classify off-specification gas as pipeline-quality, suppressing the gas diversion and blending that an off-specification classification would require. With low-Wobbe-index gas distributed without appliance adjustment, residential furnaces, water heaters, and ranges throughout the affected distribution zone operate in a rich combustion regime with insufficient combustion air: the excess fuel concentration in the burner flame produces CO at concentrations of 100–5,000 ppm above individual burners, well above the OSHA TWA of 25 ppm (ACGIH) and OSHA PEL of 50 ppm (29 CFR 1910.1000 Table Z-1), and above the immediately dangerous to life and health (IDLH) concentration of 1,200 ppm for exposures above 30 minutes. CO poisoning from gas appliance off-Wobbe combustion — because the gas is colourless and odourized (the TBM/THT odorant is perceived as the expected gas odour, not as an alarm indicator) — is particularly insidious: occupants do not recognise the symptoms (headache, dizziness, nausea, confusion) as CO poisoning until loss of consciousness. The US CPSC estimates approximately 160 residential deaths per year from CO poisoning related to gas appliances; off-specification Wobbe gas distributed undetected to a large residential service zone would multiply this consequence proportionally to the affected zone size.

3. Odorization injection rate display AI (Odotech OdoWatch AI, Brooks Instrument AI, Welker RMO odorant injection AI)

49 CFR 192.625 (Odorization of gas) requires that natural gas distributed through a distribution system be odorized sufficiently that a gas concentration in air of one-fifth (1/5) of the lower explosive limit (LEL) can be detected by a person with a normal sense of smell — a requirement that translates to an odorant (typically tert-butyl mercaptan, TBM, or tetrahydrothiophene, THT, or blends) concentration in the distributed gas of approximately 0.5–2.0 ppm (parts per million by volume) depending on the odorant’s human olfactory threshold. TBM human olfactory threshold: approximately 0.5–1.0 ppb (0.0005–0.001 ppm); THT human threshold: approximately 0.5–2.0 ppb. The 1/5 LEL methane concentration of approximately 1% by volume in air corresponds to an odorant concentration of approximately 1.0–1.4 mg/m³ (0.4–0.5 ppm TBM) at the 49 CFR 192.625 minimum level. The odorization system at the city gate station — typically a volumetric or capillary injection system that meters liquid odorant (TBM or THT at 0.5–5 gallons per million standard cubic feet of gas) into the gas stream — is monitored by inline odorant concentration analysers and injection rate displays that are rendered as SCADA screens showing current injection rate (gallons per MMSCF), downstream odorant concentration (ppm or mg/m³), and alarm status. AI systems process rendered odorization display images to classify odorant adequacy: compliant (odorant concentration at or above 49 CFR 192.625 minimum for 1/5 LEL detection), marginal (approaching minimum — injection pump inspection required), non-compliant (below 192.625 minimum — emergency injection pump restart and distribution zone odorant emergency injection required), and over-odorized (significantly above minimum — customer complaints expected; injection rate reduction required).

An adversarial perturbation on a rendered odorization display that suppresses a non-compliant odorant level — applying a ±10 DN upward shift in the pixel region encoding the digital odorant injection rate or concentration readout (raising the apparent odorant level from below the 49 CFR 192.625 minimum to an apparently adequate level) — causes the odorization monitoring AI to classify under-odorized distributed gas as compliant with the 49 CFR 192.625 minimum, suppressing the emergency injection pump restart and odorant emergency injection that a non-compliant classification would require. With under-odorized gas distributed throughout the affected zone, any gas leak from a distribution main, service line, meter connection, or appliance connector will not be detectable by human smell at the 1/5 LEL concentration required by 49 CFR 192.625: occupants of homes and buildings with active gas leaks from corroded fittings, displaced joints, or accidental mechanical damage will not detect the methane accumulation until the gas concentration approaches or exceeds the LEL (5% by volume in air) — at which point the gas-air mixture is immediately explosive on contact with any ignition source (spark, pilot light, electric switch). The 2010 San Bruno PG&E pipeline explosion — while not a distribution odorant failure — illustrated that residents in the affected zone did not detect the gas release before the explosion because the leak developed rapidly from a large-diameter pipe rupture. Under-odorization of a residential distribution zone replaces rapid detection (the odorant warning) with zero warning time, converting any distribution system leak from a detectable condition (odorant detected → evacuation → emergency repair) to an undetectable accumulation (no odorant → no evacuation → gas reaches LEL → explosion on ignition). PHMSA Advisory Bulletin ADB-10-08 emphasises that pipeline safety requires odorisation as a non-negotiable last line of defence for consumer leak detection — adversarial suppression of the odorization AI defeats this non-negotiable defence without any automated system becoming aware.

4. SCADA leak survey display AI (methane FLIR OGI survey AI — FLIR GF320 gas imaging AI, Sensirion Methane Grid AI, Heath RMLD-IS remote methane laser detector AI, Bridger Photonics LIDAR OGI AI)

49 CFR 192.706 (Transmission lines: patrolling) and 192.723 (Distribution lines: leakage surveys and records) establish mandatory leak survey intervals for natural gas distribution systems: mains in business districts require leakage surveys at intervals not exceeding 1 year; residential mains require surveys at intervals not exceeding 3 years; service lines require periodic surveys. Modern LDC leak survey programmes use methane optical gas imaging (OGI) cameras (FLIR GF320 or equivalent; infrared absorption imaging of the methane CH4 absorption band at 3.3 μm; sensitivity approximately 10 ppm·m methane column density at 10 metres range), remote methane laser detectors (RMLD; Heath Consultants RMLD-IS; tunable diode laser absorption spectroscopy at 1.65 μm methane CH4 band; range 20–50 metres), or drone-mounted methane sensors (Bridger Photonics Gas Mapping LiDAR; laser-based methane imaging from drone altitude). The leak survey instrument output is rendered as a display screen showing the current methane column density (ppm·m), alarm threshold status, geo-referenced survey track map with leak location markers, and survey history trend. AI systems process rendered OGI display images — including false-colour gas column density maps overlaid on visible-light images of the survey area, rendered RMLD concentration-distance charts, and SCADA map displays showing leak classification and priority — to classify leak survey findings: clear (no methane above background 1.9 ppm, area clear), minor indication (low methane reading above background — schedule follow-up survey), Grade 1 leak (methane above 80% LEL or confined space hazard — immediate excavation and repair required per 49 CFR 192.703), and Grade 2 leak (non-hazardous confined space leak — schedule repair within 6 months per PHMSA Distribution Integrity Management Program).

An adversarial perturbation on a rendered methane OGI display or RMLD concentration display that suppresses a leak indication — applying a ±8 DN downward shift in the pixel region encoding the methane column density colour signature (reducing the apparent methane plume colour from the alarm range — rendered in orange-red for column densities above the Grade 1 leak threshold — to the clear range rendered in blue-green for background methane levels) — causes the leak survey AI to classify an active Grade 1 or Grade 2 distribution system leak as a clear area, suppressing the excavation and repair that a Grade 1 classification requires per 49 CFR 192.703 or the scheduled repair that a Grade 2 classification requires under the PHMSA Distribution Integrity Management Program. With a Grade 1 leak undetected by the AI survey system, methane continues to accumulate in the soil surrounding the leak point and to migrate through permeable soil layers toward basements, crawlspaces, utility vaults, and other below-grade enclosed spaces. Methane accumulation in an enclosed below-grade space — at the surface expression point of a distribution system leak — can reach the LEL (5% by volume) within 24–72 hours of leak initiation depending on the leak flow rate and the enclosed space volume; at LEL in a basement or crawlspace, any ignition source (electric appliance, telephone, light switch) can trigger an explosion. Under PHMSA Mandatory Incident Reporting (49 CFR 191.15), incidents involving gas distribution systems that result in injuries, deaths, or property damage above $50,000 must be reported to PHMSA within 30 days; adversarial suppression of the leak survey AI converts an active reportable-incident-precursor Grade 1 leak into an unrecorded survey result, eliminating the PHMSA safety data trail that normally triggers regulatory follow-up on distribution systems with repeated Grade 1 leak indications. Methane also carries a global warming potency of approximately 80× CO2 over a 20-year horizon (IPCC AR6, GWP20 for CH4): undetected distribution system methane leaks that exceed EPA GHG Reporting Program Subpart W (40 CFR Part 98 Subpart W) reporting thresholds — applicable to natural gas distribution operators with combined leak emissions above 25,000 metric tonnes CO2e/yr — constitute a Subpart W non-compliance if AI leak survey suppression prevents the required leak survey data from reaching the annual emission report.

Integration: urban gas distribution city gate station AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for urban gas distribution city gate station AI belongs at every rendered-image ingestion boundary in the gas distribution monitoring and control pipeline — before pressure regulator downstream display AI processes rendered SCADA or local gauge images, before gas chromatograph Wobbe index and Btu display AI processes rendered GC screen images, before odorization injection rate display AI processes rendered odorant injection system images, and before methane OGI or RMLD SCADA leak survey AI processes rendered OGI or RMLD concentration displays. Threshold 30 for urban gas distribution AI contexts reflects the consequence envelope of residential over-pressurization and house explosion (Columbia Gas 2018: 1 killed, 22 injured, 8,600 customers, $143M consequence; NiSource criminal guilty plea), CO poisoning from off-Wobbe residential appliance combustion, and undetectable gas leak accumulation from under-odorized distribution below the 49 CFR 192.625 minimum.

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"

# Urban gas distribution AI contexts: threshold 30
# 49 CFR Part 192 (PHMSA minimum federal safety standards for gas pipelines);
# AGA Report 3/7/12 (gas measurement); ASME B31.8 (gas piping systems);
# EPA 40 CFR Part 98 Subpart W (GHG reporting for gas distribution).
GAS_DISTRIBUTION_THRESHOLD = 30


class GasDistributionAIContext(Enum):
    PRESSURE_REGULATOR  = "pressure_regulator"  # Downstream pressure display AI
    WOBBE_INDEX_GC      = "wobbe_index_gc"       # GC Wobbe index display AI
    ODORIZATION_RATE    = "odorization_rate"     # Odorant injection rate display AI
    LEAK_SURVEY         = "leak_survey"          # OGI/RMLD SCADA leak survey AI


class AdversarialGasDistributionImageError(Exception):
    """Raised when Glyphward detects adversarial content in a gas
    distribution city gate AI rendered image above threshold 30.

    Consequence if not raised:
    - PRESSURE_REGULATOR: over-pressurization suppressed → residential
      distribution above MAOP → appliance bypass → gas accumulation →
      house explosion; Columbia Gas 2018 mechanism (1 killed, 22 injured,
      8,600 customers, $143M consequence).
    - WOBBE_INDEX_GC: off-spec Wobbe suppressed → low-Wobbe gas to
      residential → rich combustion → CO >1,200 ppm IDLH above burners
      → CO poisoning; or high-Wobbe → flashback → appliance fire.
    - ODORIZATION_RATE: under-odorization suppressed → gas below
      49 CFR 192.625 minimum odorant → undetectable gas leak →
      LEL accumulation in enclosed space → explosion on ignition.
    - LEAK_SURVEY: Grade 1 leak suppressed → methane accumulation in
      basement/crawlspace → LEL → explosion; PHMSA §192.703 non-compliance.
    Fail-safe: halt gas distribution AI classification; require manual
    pressure gauging / GC verification / odorant sampling / physical
    leak survey before resuming AI-driven gas management.
    """

    def __init__(self, scan_id: str, score: int,
                 context: GasDistributionAIContext,
                 ldc_id: str, station_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.ldc_id = ldc_id
        self.station_id = station_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial gas distribution image: "
            f"context={context.value} score={score} "
            f"ldc={ldc_id} station={station_id} scan_id={scan_id}"
        )


async def scan_gas_distribution_image(
    image_bytes: bytes,
    context: GasDistributionAIContext,
    ldc_id: str,
    station_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a gas distribution city gate AI rendered image for adversarial content.

    Fail-safe contract: AdversarialGasDistributionImageError or httpx error →
    halt gas distribution AI classification; require manual pressure gauge
    reading (PRESSURE_REGULATOR), GC grab sample (WOBBE_INDEX_GC), odorant
    sniff test per 49 CFR 192.625 (ODORIZATION_RATE), or bar-hole survey
    (LEAK_SURVEY) before resuming AI-driven gas management decisions.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"gas_distribution:{context.value}:{ldc_id}:{station_id}",
        "metadata": {
            "ldc_id": ldc_id,
            "station_id": station_id,
            "context": context.value,
            "image_sha256": image_hash,
        },
    }
    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"] > GAS_DISTRIBUTION_THRESHOLD:
        raise AdversarialGasDistributionImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            ldc_id=ldc_id,
            station_id=station_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_gas_distribution_image at each gas distribution city gate station AI rendered-image ingestion boundary: before pressure regulator downstream display AI (threshold 30), before GC Wobbe index display AI (threshold 30), before odorization injection rate display AI (threshold 30), and before methane OGI/RMLD SCADA leak survey AI (threshold 30). On AdversarialGasDistributionImageError for PRESSURE_REGULATOR context: immediately close the distribution zone isolation valve, initiate manual downstream pressure measurement at the nearest district pressure check tap, and notify PHMSA under 49 CFR 191.22 (gas distribution system emergency reporting) if over-pressurization has already reached customer service lines before resuming AI-driven pressure management. See also: pipeline integrity inspection AI prompt injection (related gas transmission pipeline AI adversarial surface) and smart grid and power distribution AI prompt injection (related utility infrastructure AI adversarial injection context). Get early access

Related questions