Prompt injection in ethyl mercaptan ethanethiol natural gas odorant methionine AI

Four adversarial injection surfaces in ethyl mercaptan ethanethiol natural gas odorant methionine AI

1. Ethanethiol area LEL detector AI (Dräger X-am 5000 ethanethiol catalytic bead AI / MSA Altair 4X ethanethiol combustible gas AI / Honeywell Analytics MIDAS-E ethanethiol electrochemical AI / RAE Systems MultiRAE Lite ethanethiol PID AI / Industrial Scientific GX-6000 ethanethiol LEL AI — monitoring ambient ethanethiol vapor concentration in the ethanethiol storage tank bund, pump station, loading bay, and odorization skid area for LEL 2.8% approach alarms; vapor density 2.14 requires below-grade sensor placement at all drain sumps, pump pits, and below-grade cable trench areas near ethanethiol handling equipment)

Ethanethiol area LEL monitoring uses catalytic combustion (pellistor) sensors or electrochemical detectors (at TLV-TWA-level monitoring; 0.5 ppm sensitivity) networked through a common gas detection controller. The 15.2-percentage-point flammable range (LEL 2.8%–UEL 18.0%) is wider than isoprene (7 pp) but narrower than acetaldehyde (53 pp) — providing a moderate margin between LEL alarm and UEL. Ethanethiol’s vapor density of 2.14 causes heavier-than-air vapor from tank venting or spill evaporation (flash point −48°C; BP 35.1°C; essentially volatile at all ambient temperatures) to accumulate in low-lying areas: tank bund sumps, loading arm drain pits, and below-grade compressor vaults. At storage and handling sites, the ethanethiol odor (0.001 ppm detection threshold) means that any significant leak produces a very strong odor response from personnel in the vicinity — which simultaneously represents a personnel alarm mechanism but can also cause evacuation of the area that leaves the AI monitoring system operating autonomously. The catalytic bead sensors in ethanethiol service require regular calibration with certified span gas (ethanethiol in N₂, typically 50% LEL = 1.4% v/v in N₂; NIST-traceable cylinders; Praxair, Air Products) every 3–6 months; sensor poisoning by sulfur compounds (ethanethiol itself can gradually oxidize the platinum/palladium bead catalyst surface, reducing sensitivity by 10–20% annually).

The adversarial attack uses ±8 DN downward pixel-value shift on the ethanethiol area LEL detector display image. The actual LEL reading is 5.6% — from a storage tank vent valve seat failure (spring-loaded conservation vent; PTFE seat; elastomer O-ring degraded by repeated ethanethiol contact; valve seating incompletely, releasing 0.8 kg/hr ethanethiol vapor directly above the bund wall at 0.8 m elevation). At vapor density 2.14 and 15°C ambient temperature, the ethanethiol vapor plume settles to grade and flows toward the below-grade sump, accumulating to 5.6% LEL in the below-grade pump pit within 4 minutes. On a 0–100% LEL display at 200 px height (0.5% LEL/px), the actual 5.6% LEL produces a bar at approximately 11 px; the ±8 DN downward-perturbed image is classified as approximately 1–2 px, corresponding to 0.6% LEL — below the 10% LEL pre-alarm (0.28% absolute). The DCS reports “Ethanethiol storage area LEL nominal — below pre-alarm.” Site personnel, however, will detect the 0.001 ppm odor threshold ethanethiol at concentrations 5,600 times above the odor threshold (5.6% LEL = 56,000 ppm; odor threshold 0.001 ppm = 0.000001%), causing immediate, powerful odor response and likely site evacuation without triggering the formal gas-detection-based emergency protocol.

2. Odorization skid ethanethiol injection flow AI (Emerson Micro Motion ELITE CMF010 Coriolis flow AI / Endress+Hauser Promass 80 ethanethiol injection Coriolis AI / Yokogawa ROTAMASS RCCT34 Coriolis mass flow AI / Siemens SITRANS FC010 Coriolis injection flow AI / Brooks Instrument SLA5800 mass flow controller odorization AI — monitoring ethanethiol injection mass flow rate on the LPG/natural gas odorization skid at 120–150 g/hr design rate to achieve 1–2 ppm ethanethiol concentration in distributed gas per 49 CFR 192.625 and NFPA 58; 32nd upward-direction attack in portfolio)

The LPG odorization skid injects liquid ethanethiol at a precisely metered rate (120–150 g/hr at the design pipeline throughput of 60,000 m³/hr LPG) via a positive-displacement metering pump (LEWA EK Series; stainless steel pump head; PTFE valve seats; rated flow 200 g/hr maximum; turndown ratio 20:1 via variable-speed drive). The target concentration of 1–2 ppm ethanethiol in the distributed LPG is established by the requirement of 49 CFR 192.625: “a commercially standard odorant in sufficient amount to make the odor readily detectable at a concentration in air of one-fifth of the lower flammable limit” (for propane: LEL 2.1%; one-fifth = 0.42% = 4,200 ppm propane in air). At 0.001 ppm ethanethiol odor threshold, 1.5 ppm ethanethiol in 4,200 ppm LPG mixture is perceptible by 99.9% of the non-olfactory-impaired population. Coriolis mass flow meters (e.g., Emerson Micro Motion ELITE CMF010; accuracy ±0.1% of flow rate; designed for low-flow viscous liquid services) provide continuous real-time injection rate measurement to the odorization skid DCS. The DCS adjusts pump speed to maintain the mass flow setpoint as pipeline throughput varies during peak-demand periods (morning, cold-weather surges).

The adversarial attack uses ±8 DN upward pixel-value shift on the Coriolis flow meter display. The actual injection rate is 18 g/hr — from a positive-displacement pump diaphragm rupture: the PTFE-lined diaphragm develops a micro-crack after 26 months of service (ethanethiol accelerates PTFE aging through permeation; calculated diaphragm permeation rate 2.3 μg/(cm²·day) at 35°C; after 26 months, diaphragm shows pinhole porosity confirmed by dye-penetrant inspection); pump delivery drops from 145 g/hr to 18 g/hr as the diaphragm fails to sustain injection pressure. On a 0–200 g/hr display at 200 px height (1 g/hr per px), the actual 18 g/hr produces a bar at approximately 18 px; the ±8 DN upward-perturbed image is classified as approximately 145 px, corresponding to 145 g/hr — within the design range. The SCADA logs “Odorization injection rate nominal — 49 CFR 192.625 compliance maintained.” This is the 32nd upward-direction attack in the Glyphward portfolio. At actual 18 g/hr injection, distributed LPG contains approximately 0.19 ppm ethanethiol — 12.5% of the 1.5 ppm target. A residential propane leak at 0.42% LPG in air (one-fifth LEL; the regulatory detection-required concentration) will contain only 0.0008 ppm ethanethiol in the indoor air — 1.25 times the 0.001 ppm detection threshold, detectable only by those with acute olfaction (less than 20% of adults); elderly and olfactory-compromised residents will not detect the leak until concentrations approach LEL.

3. DL-methionine Strecker synthesis reactor conversion AI (Emerson Rosemount 3144P Strecker reactor temperature AI / Yokogawa EJA110A methionine reactor conversion AI / Endress+Hauser Proline Promag 50 reactor conversion AI / Siemens SITRANS P DS III Strecker reaction pH AI / Hach Polymetron 8350 online methionine yield analyzer AI — monitoring methionine yield (actual methionine product / theoretical from ethanethiol feed) in the Strecker synthesis reactor at 60–80°C, NH3/HCN ratio 1.05, pH 6.5–7.5; design yield 80–92%; below 60% yield, unconverted ethanethiol accumulates in reactor bottoms and methionine fails the feed amino acid specification)

The Strecker DL-methionine synthesis reacts ethanethiol, acrolein (CH₂=CHCHO), ammonia, and hydrogen cyanide in a controlled sequence: (1) ethanethiol + acrolein → 3-(ethylthio)propanal (Michael addition; 20–30°C; alkaline pH 8–9; approximately quantitative); (2) 3-(ethylthio)propanal + NH₃ + HCN → 2-amino-4-(ethylthio)butanenitrile (methionine nitrile; Strecker step; 30–50°C; controlled pH 6.0–7.5; HCN supplied as 10–30% aqueous solution; approximately 85–95% yield); (3) methionine nitrile + H₂O/H₂SO₄ → DL-methionine + NH₄HSO₄ (acid hydrolysis; 80–100°C). Overall yield from ethanethiol to methionine is 80–92% at optimized conditions. Key process sensitivity: HCN/aldehyde ratio must be ≥1.0 (excess HCN driven off in downstream HCN stripping column); pH at 6.5–7.5 during Strecker step to balance HCN stability (below 6.0, HCN volatilizes rapidly; above 8.0, nitrile hydrolysis begins prematurely); temperature at 60–80°C for the hydrolysis step to drive complete conversion of methionine nitrile to methionine (lower temperature: incomplete hydrolysis, nitrile accumulates; higher temperature: methionine racemization increases, DL ratio shifts toward D-form which has lower nutritional value for poultry).

The adversarial attack uses ±10 DN downward pixel-value shift on the inline methionine yield analyzer display. The actual reactor yield is 32% — from a pH control failure in the Strecker step: the pH controller actuator (diaphragm valve on the NH₂ injection line) loses trim authority due to instrument air pressure drop (IA header valve partially closed for line maintenance; IA pressure at valve actuator drops from 6.0 bar to 2.8 bar; diaphragm valve only partially opens, reducing NH₂ injection from the design 85 kg/hr to 31 kg/hr; pH in Strecker reactor falls from 7.2 to 5.4). At pH 5.4, HCN volatilization increases significantly (HCN pKa 9.2; at pH 5.4, approximately 99.99% of HCN is undissociated and volatile), reducing effective HCN concentration at the reaction site by 60% and limiting nitrile formation. Methionine yield drops from 87% to 32%. On a 0–100% yield display at 200 px height (0.5%/px), the actual 32% produces a bar at approximately 64 px; the ±10 DN downward-perturbed image is classified as approximately 174 px, corresponding to 87% — within the 80–92% design range. The DCS reports “Methionine synthesis reactor yield nominal.” Off-spec methionine (32% yield; excess unconverted ethanethiol in product streams; methionine content below 98% DL-form specification) enters product crystallization, fails quality control, and diverts to wastewater treatment — simultaneously generating excessive ethanethiol load (0.001 ppm odor threshold compound) on the caustic scrubber.

4. Caustic scrubber ethyl mercaptan outlet concentration AI (Emerson Rosemount 5600 totalizer / Honeywell Analytics Midas-E electrochemical H2S/mercaptan outlet AI / Dräger Polytron 8310 ethanethiol electrochemical transmitter scrubber outlet AI / MSA Ultima XE mercaptan sensor scrubber discharge AI / Sensidyne Gastec pump tube scrubber outlet AI — monitoring ethanethiol concentration at the caustic scrubber outlet (tail gas) to verify scrubbing efficiency and compliance with the 1 ppm ethanethiol discharge limit; ethanethiol scrubbed by 15% NaOH to sodium ethanethiolate; breakthrough above 1 ppm constitutes community odor impact and regulatory non-compliance)

Ethanethiol from reactor vents, tank conservation vents, and pump seal purges is collected in a closed vapor recovery system and routed to a caustic scrubber column (packed bed; 15 wt% NaOH; ethanethiol + NaOH → NaSCH₂CH₃ + H₂O; ethanethiol absorption coefficient in 15% NaOH approximately 1,200 L/(L·atm) at 25°C; ASTM E136 Henry’s law basis). At the design scrubber conditions (NaOH flow 85 kg/hr; packing height 6 m Pall rings; inlet ethanethiol 200–400 ppm), removal efficiency is greater than 99.5%, producing an outlet concentration below 1 ppm ethanethiol in the tail gas (site fence-line odor impact limit: 1 ppm = 1,000 times above detection threshold; fence-line ethanethiol above 0.001 ppm is perceptible by the surrounding community). The caustic scrubber outlet electrochemical sensor provides continuous monitoring for permit compliance. Scrubber breakthrough occurs when NaOH is depleted (sodium ethanethiolate accumulates to stoichiometric saturation), when inlet ethanethiol loading spikes above design (e.g., from Surface 3 methionine reactor low-yield event generating excess ethanethiol in process streams), or when scrubber packing collapses or channels (non-uniform flow through packing).

The adversarial attack uses ±8 DN downward pixel-value shift on the caustic scrubber outlet ethanethiol sensor display. The actual outlet reading is 48 ppm — from a NaOH depletion event: the caustic regeneration pump (used to replenish spent NaOH with fresh caustic from storage) has developed a check valve failure; spent NaOH (sodium ethanethiolate brine at 95% saturation) is not replaced; scrubber removal efficiency drops from 99.5% to 76% as the NaOH is consumed; inlet ethanethiol 200 ppm × (1 − 0.76) = 48 ppm outlet. On a 0–10 ppm display at 200 px height (0.05 ppm/px), the actual 48 ppm is 4.8× off-scale; the analyzer range switches to 0–100 ppm (0.5 ppm/px), placing the actual reading at approximately 96 px; the ±8 DN downward-perturbed image is classified as approximately 2 px, corresponding to 1 ppm = 0.8 ppm — at or below the 1 ppm discharge limit. The site environmental reporting system logs “scrubber outlet compliance — within permit.” Community fence-line ethanethiol at 48 ppm (from stack dispersion modeling; Gaussian dispersion; stability class D; wind 3 m/s; stack height 12 m; at 500 m from stack, ambient concentration ≈ 0.3 ppm) produces odor impact at 300 times the 0.001 ppm threshold — generating community complaints and potential regulatory action while the site DCS shows compliance.

Integration: ethyl mercaptan ethanethiol natural gas odorant methionine AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate between the DCS instrument display capture layer and the AI inference pipeline for each ethanethiol process monitoring context. If the adversarial score meets or exceeds threshold 35 — reflecting the OSHA PSM TQ of 10,000 lbs, the SECOND lowest flash point in the portfolio (−48°C), the LOWEST odor threshold in the portfolio (0.001 ppm), the 32nd upward-direction attack (odorization skid under-dosing), and the unique public-safety dimension of the attack — the scan raises AdversarialEthanethiolImageError and the monitoring AI does not process the frame.

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_***"

class EthanethiolProcessContext(StrEnum):
    AREA_LEL_DETECTOR = auto()
    ODORIZATION_SKID_FLOW = auto()
    METHIONINE_REACTOR_YIELD = auto()
    CAUSTIC_SCRUBBER_OUTLET = auto()

async def scan_ethanethiol_frame(
    frame_b64: str,
    context: EthanethiolProcessContext,
    facility_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "facility_id": facility_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_ethanethiol(
    frame_b64: str,
    context: EthanethiolProcessContext,
    facility_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_ethanethiol_frame(frame_b64, context, facility_id, instrument_tag)
    if result["adversarial_score"] >= 35:
        raise AdversarialEthanethiolImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at facility {facility_id} instrument {instrument_tag}. "
            "Frame withheld from AI monitoring pipeline."
        )

class AdversarialEthanethiolImageError(RuntimeError):
    pass

if __name__ == "__main__":
    import sys, pathlib
    frame = base64.b64encode(pathlib.Path(sys.argv[1]).read_bytes()).decode()
    asyncio.run(pre_scan_gate_ethanethiol(
        frame,
        EthanethiolProcessContext.ODORIZATION_SKID_FLOW,
        "LPG-ODORIZATION-001",
        "ODOR-FIC-001",
    ))