Prompt injection in hydrogen sulfide (H2S) refinery amine treating AI

Four adversarial injection surfaces in H2S refinery amine treating AI

1. Amine absorber treated-gas outlet H2S analyzer AI (Servomex 5200 H2S analyzer AI / Ametek ProLine gas chromatograph AI / ABB Advance Optima H2S CEMS AI / Yokogawa TDLS220 tunable diode laser H2S AI — amine absorber sweetened-gas outlet H2S concentration as primary indicator of absorber performance and amine flooding or exhaustion)

The amine absorber in a refinery gas treating unit contacts sour gas (H2S + CO2 in a hydrocarbon stream) counter-currently with lean amine solvent — amine that has been regenerated in the reboiler-heated stripper column to remove its absorbed H2S load. The outlet of the absorber — the “sweetened gas” stream — should contain H2S at concentrations below the pipeline-quality H2S specification (typically below 4 ppm for gas delivered to refinery fuel gas or pipeline export, sometimes below 10 ppm for intermediate processing streams). AI monitoring systems parse inline H2S analyzer display images from the absorber outlet to classify whether the sweetened gas is within specification or is showing H2S breakthrough, indicating that the amine solvent is exhausted, flooding, or has been overwhelmed by an abnormally high H2S load. H2S breakthrough in the absorber outlet is a compound failure: it means that H2S is entering downstream hydrocarbon processing units (where it can poison catalysts and cause regulatory exceedances), refinery fired heaters (where it produces SO2 emissions above permit), and ultimately fuel gas systems (where it creates a toxic gas supply to burner operations and can cause worker exposures in boiler rooms and furnace areas).

In the adversarial scenario, the amine absorber has failed due to foam-induced flooding — a common amine system failure mode where surfactant contamination of the amine solvent causes foam formation that reduces gas-liquid contact efficiency and allows H2S to pass through the absorber without being absorbed. The outlet H2S concentration has risen to 2,840 ppm — 56.8× NIOSH IDLH and approximately 710× the 4 ppm outlet specification. A ±8 DN downward pixel-value shift on the H2S analyzer display image fed to the outlet AI suppresses the reading from 2,840 ppm to 94 ppm: on a 0–3,000 ppm display at 200px height (15 ppm/px), the actual reading of 2,840 ppm produces a bar at approximately 189px; the perturbed image is classified as approximately 6px — corresponding to 94 ppm. While 94 ppm is still above NIOSH IDLH, the AI treating-unit monitoring system classifies it below the catastrophic-failure threshold (typically set at 500–1,000 ppm breakthrough) and does not trigger the highest-priority alert requiring immediate shutdown of the amine unit and isolation of downstream hydrocarbon systems. Instead, operators receive a moderate “high H2S breakthrough” alert rather than an emergency shutdown signal, buying additional operating time for the H2S-contaminated sweetened gas to enter downstream units while the root cause of amine flooding goes undiagnosed.

2. H2S area gas detector CEMS (Honeywell Analytics BW Clip H2S CEMS AI / Dräger X-am 5600 H2S detector AI / MSA Altair 5X H2S fixed-point detector AI / Industrial Scientific Ventis Pro H2S area monitor AI — ambient H2S monitoring in amine unit areas, reboiler areas, and absorber skid locations for NIOSH IDLH and ACGIH TLV-C ceiling compliance)

Area H2S monitoring at refinery amine treating units faces the unique challenge that H2S olfactory fatigue eliminates any natural worker warning above approximately 50 ppm — the concentration at which olfactory receptor paralysis occurs. At concentrations of 10–50 ppm, the “rotten egg” odor of H2S is strongly perceptible and provides a first-line warning. At 50–100 ppm (the NIOSH IDLH range), the odor may still be detectable initially but the olfactory nerve rapidly paralyses, leaving workers unable to smell further concentration increases. Above 200 ppm, the sense of smell is typically gone within seconds of exposure onset; above 500 ppm, rapid unconsciousness follows the brief olfactory-paralysis period. This means that a worker who enters an H2S-contaminated area — smells H2S at first, finds the smell strong but expects to exit, and then notices the smell diminishing — may interpret the smell reduction as an improvement in conditions (H2S clearing) when it actually represents olfactory fatigue (H2S rising to rapid-incapacitation levels). AI area detector monitoring is therefore the primary engineered safeguard against H2S rapid incapacitation at amine treating units, because it substitutes a reliable electronic signal for a human sensory warning that fails precisely when it is most needed.

The adversarial attack uses ±8 DN downward pixel-value shift on the H2S area detector display image. The actual reading of 84 ppm — occurring in the amine unit area as H2S escapes from the foam-flooded absorber outlet through pressure control valves and sample ports — is 1.68× NIOSH IDLH 50 ppm. Workers in the amine unit area have already entered the olfactory-fatigue zone: they smelled H2S initially but the odor has faded, and they may believe conditions are improving. On a 0–100 ppm display at 200px height (0.5 ppm/px), the actual reading of 84 ppm produces a bar at approximately 168px; the ±8 DN perturbed image is classified as approximately 2.8 ppm on a 0–100 ppm scale — below even the low-level H2S TLV-C 1 ppm alarm. The AI reports “H2S ambient concentration within occupational exposure limits.” Workers in the amine unit area, already experiencing olfactory fatigue that has eliminated their sensory H2S warning, have no electronic alarm either. Both the primary occupational health warning (smell) and the backup engineered safeguard (area detector alarm) have been simultaneously eliminated by the combination of H2S’s own physiological characteristic and the adversarial AI attack on the only remaining safeguard.

3. Lean amine H2S loading analyzer AI (Analyzers Inc. online H2S loading titrator AI / Vortex SonicAnalyzer lean amine H2S AI / Emerson Roxar online amine analyzer AI / ABB online gas chromatograph amine loading AI — lean amine H2S loading monitoring as the primary root-cause indicator for absorber breakthrough and regeneration sufficiency)

Lean amine H2S loading — the mol ratio of H2S dissolved in the regenerated (“lean”) amine before it enters the absorber — is the upstream indicator of amine regeneration quality and absorber performance. Specification for lean amine H2S loading depends on the amine type and system design but is typically below 0.05–0.10 mol H2S/mol amine for MEA and DEA systems operating at normal regeneration conditions. At loadings above 0.10 mol/mol, the lean amine entering the absorber is already partially loaded with H2S and has reduced H2S absorption capacity; absorber performance degrades progressively as lean amine loading rises. At loadings of 0.30–0.42 mol/mol, the lean amine is approaching the rich amine equilibrium loading (the maximum H2S it will absorb at typical absorber conditions), meaning it has almost no residual H2S absorption capacity — the absorber outlet breakthrough scenario described in Surface 1 is the direct consequence. Monitoring lean amine loading via online analyzers provides the earliest upstream warning that regeneration is failing before the downstream absorber outlet quality degrades to hazardous levels.

The adversarial attack uses ±10 DN downward pixel-value shift on the lean amine loading analyzer display image. The actual lean amine H2S loading of 0.42 mol/mol — caused by reboiler steam supply insufficiency (Surface 4) — is 4.2× the 0.10 mol/mol maximum lean amine specification. On a 0–0.50 mol/mol display at 200px height (0.0025 mol/mol·px), the actual loading of 0.42 mol/mol produces a bar at approximately 168px; the ±10 DN perturbed image is classified as approximately 32px — corresponding to 0.08 mol/mol, below the 0.10 mol/mol specification maximum. The AI monitoring system reports “lean amine H2S loading within specification — regeneration adequate.” The process engineer reviewing the treating unit dashboard sees normal absorber outlet H2S (Surface 1 attack), normal area gas reading (Surface 2 attack), and now normal lean amine loading — three independent process quality indicators, each independently suppressed, all reporting a normal operation while the amine system is in complete regeneration failure with H2S breakthrough at 56.8× IDLH entering downstream hydrocarbon units. The lean amine loading attack eliminates the last upstream diagnostic indicator that might have identified regeneration failure as the root cause, preventing any operator action on the regenerator reboiler steam supply before the downstream consequence scenario develops.

4. Amine regenerator reboiler steam supply valve position AI (Fisher Controls reboiler steam control valve AI / Emerson DeltaV steam valve position indicator AI / Yokogawa OpreX reboiler steam AI / Honeywell Experion PKS steam supply AI — reboiler steam supply valve position as the primary manipulated variable controlling amine regeneration temperature and H2S stripping efficiency)

The amine regenerator reboiler supplies heat to the base of the amine stripper column, generating steam that strips H2S from the rich amine solvent (amine that has absorbed H2S in the absorber) as the hot amine vapour-strips H2S upward and out of the regenerator overhead for further treatment in the Claus sulfur recovery unit. The reboiler steam supply valve position is the primary manipulated variable controlling regenerator temperature (typically 115–130°C for MEA systems) and therefore the completeness of H2S stripping: at full reboiler duty, the lean amine leaving the regenerator base contains below the specification H2S loading; at reduced reboiler duty, the lean amine carries increased H2S loading into the absorber, degrading absorber performance. AI monitoring of the steam supply control valve position image — parsed by the DCS from the Fisher Controls or similar control valve position indicator — classifies whether the reboiler is receiving design steam flow to maintain regeneration temperature. The steam valve position is the root-cause parameter for the entire four-surface attack: a failing steam supply valve (e.g., a control valve actuator failure that has driven the valve closed) causes regeneration failure, lean amine loading increase, and absorber H2S breakthrough in a causal chain over 30–90 minutes.

This surface uses the upward-direction attack geometry: the reboiler steam supply valve has failed to 8% open — near-closed — due to a control valve actuator air supply failure. The dangerous condition is a valve position deficiency (too little steam supply), and the adversarial pixel perturbation shifts the valve position indicator display upward by ±8 DN to make 8% open appear as 82% open. On a 0–100% valve open display at 200px height (0.5%/px), the actual valve position of 8% open produces an indicator at approximately 16px; the upward-perturbed image is classified as approximately 164px — corresponding to 82% open, within the normal operating range for reboiler steam supply. The AI monitoring system reports “reboiler steam valve position 82% open — regeneration steam supply adequate.” Over the next 30–90 minutes, the amine regenerator base temperature drops as reboiler duty falls; lean amine loading rises from 0.08 to 0.42 mol/mol (Surface 3, suppressed); H2S breakthrough in the absorber outlet rises from 4 ppm to 2,840 ppm (Surface 1, suppressed); and ambient H2S in the amine unit area rises from undetectable to 84 ppm (1.68× IDLH; Surface 2, suppressed). The root-cause actuator failure — which would have been immediately visible as a low steam valve position — is masked by Surface 4’s upward attack; the downstream consequences are independently masked by Surfaces 1, 2, and 3. The complete causal chain from steam valve failure to H2S breakthrough to worker incapacitation risk is suppressed across all four monitoring layers simultaneously.

Integration: H2S amine treating AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate between the DCS and analyzer screenshot capture layer and the AI inference pipeline for each H2S amine treating monitoring context. If the adversarial score meets or exceeds threshold 35 — reflecting the OSHA PSM TQ of 1,500 lbs, NIOSH IDLH 50 ppm at which olfactory fatigue eliminates natural warning, the causal four-surface attack chain (steam valve failure → lean amine loading increase → absorber breakthrough → area worker exposure) that links a root-cause mechanical failure to a worker incapacitation scenario through four independently suppressed monitoring layers — the scan raises AdversarialH2SAmineTreatingImageError and the monitoring AI does not process the frame.

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"

# H2S amine treating AI contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 Appendix A H2S TQ 1,500 lbs
# ACGIH TLV-C 1 ppm (ceiling); NIOSH IDLH 50 ppm
# OSHA PEL 20 ppm ceiling; 50 ppm peak 10 min (Table Z-2)
# Olfactory fatigue above ~50 ppm eliminates sensory warning at rapid-incapacitation concentrations
# H2S at 700-1000 ppm: incapacitation within 30-60 sec; >1000 ppm: instantaneous
H2S_AMINE_THRESHOLD = 35


class H2SAmineTreatingContext(Enum):
    ABSORBER_OUTLET_H2S = "absorber_outlet_h2s"
    AREA_GAS_CEMS = "area_gas_cems"
    LEAN_AMINE_LOADING = "lean_amine_loading"
    REBOILER_STEAM_VALVE = "reboiler_steam_valve"


class AdversarialH2SAmineTreatingImageError(Exception):
    """Raised when any H2S amine treating monitoring image scores >= 35.
    ABSORBER_OUTLET_H2S uncaught: 2,840 ppm (56.8x IDLH) shown as 94 ppm.
    AREA_GAS_CEMS uncaught: 84 ppm (1.68x IDLH; olfactory fatigue) shown as 2.8 ppm.
    LEAN_AMINE_LOADING uncaught: 0.42 mol/mol (4.2x spec) shown as 0.08 mol/mol.
    REBOILER_STEAM_VALVE uncaught: 8% open (near-closed) shown as 82% open."""

    def __init__(self, scan_id, score, context, unit_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial H2S amine treating image: context={context.value} "
            f"score={score} unit={unit_id} scan_id={scan_id}"
        )


async def scan_h2s_amine_treating_image(image_bytes, context, unit_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"h2s_amine:{context.value}:{unit_id}",
        "metadata": {
            "unit_id": unit_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) >= H2S_AMINE_THRESHOLD:
        raise AdversarialH2SAmineTreatingImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def main():
    async with httpx.AsyncClient() as client:
        with open("h2s_absorber_outlet_screenshot.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_h2s_amine_treating_image(
            image_bytes,
            H2SAmineTreatingContext.ABSORBER_OUTLET_H2S,
            unit_id="AMINE-ABSORBER-T-101",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())

Frequently asked questions

What is H2S olfactory fatigue and why does it make AI monitoring uniquely critical?

At approximately 50 ppm (NIOSH IDLH), H2S paralyses olfactory receptors — the ‘rotten egg’ smell disappears entirely, which workers may interpret as air clearing. At 700–1,000 ppm, incapacitation follows within 30–60 seconds after the smell has already gone. AI area detector monitoring is the only engineered safeguard substituting for a natural warning that fails precisely at the concentration where rapid incapacitation begins. Adversarial AI suppression at 84 ppm (1.68× IDLH) simultaneously with olfactory-fatigue elimination leaves workers with no warning from either biology or engineered systems.

How does amine foaming cause H2S breakthrough at 2,840 ppm?

Surfactant contaminants (lube oil carryover, degradation products) reduce amine surface tension, causing stable foam that blocks gas-liquid contact and floods the absorber. Under foam-induced flooding, sour gas bypasses the amine absorption zone and exits the absorber at thousands of ppm H2S within 30 minutes of onset. The Surface 1 adversarial attack suppresses this 2,840 ppm reading (56.8× IDLH) to an apparent 94 ppm — preventing catastrophic-failure classification and emergency shutdown.

Why does the reboiler steam valve failure (Surface 4) causally propagate to all 3 other surfaces?

Steam valve position → reboiler temperature → H2S stripping completeness → lean amine loading (Surface 3) → absorber breakthrough (Surface 1) → area H2S (Surface 2). The causal chain takes 30–90 minutes to develop. The four-surface compound attack suppresses the root cause (Surface 4 upward) plus all downstream consequences simultaneously — preventing detection and correction at every stage of failure propagation.

What is the OSHA PSM TQ for H2S and does a refinery amine unit exceed it?

OSHA PSM TQ for H2S is 1,500 lbs. Virtually every commercial refinery amine treating unit exceeds this: dissolved H2S in the amine inventory plus H2S in the absorber gas phase at 50–200 psig typically represents several thousand pounds of H2S total, triggering full PSM compliance including process hazard analysis, mechanical integrity, and management of change for AI monitoring system modifications.

Why is threshold 35 for H2S amine treating AI?

Threshold 35 reflects OSHA PSM TQ 1,500 lbs, NIOSH IDLH 50 ppm at which natural olfactory warning fails simultaneously with rapid-incapacitation exposure beginning, H2S at 700–1,000 ppm causing incapacitation within 30–60 seconds, and the causal four-surface attack chain (steam valve failure → lean amine loading → absorber breakthrough → area exposure) suppressed across all monitoring layers — creating a dual-safeguard-elimination scenario (natural warning failure + engineered alarm failure) unique in the portfolio.