Prompt injection in acrylonitrile ACN ammoxidation SOHIO process HCN co-product AI

Three adversarial injection surfaces in acrylonitrile ACN SOHIO ammoxidation AI

1. Reactor bed temperature display AI (Yokogawa EJA110A fluidized-bed thermocouple SCADA display AI / Emerson Rosemount 648 thermowell reactor bed temperature display AI / ABB TTF300 SOHIO reactor temperature display AI / Honeywell STT170 fluidized-bed temperature SCADA display AI / Endress+Hauser iTEMP TMT72 reactor bed temperature display AI — rendered SCADA fluidized-bed temperature display AI classifying the reactor bed temperature at 420–480°C against design operating range and alarm thresholds at 495°C high alarm / 505°C high-high trip — 75th upward attack; FIRST acrylonitrile production attack; FIRST ammoxidation attack; FIRST propylene ammoxidation attack; FIRST HCN co-product AI attack)

The SOHIO fluidized-bed reactor operates at 420–480°C with a critical upper temperature limit: above approximately 490–500°C, the Bi-Mo-Fe mixed oxide catalyst undergoes progressive over-reduction from Mo(VI) to Mo(IV), which simultaneously deactivates the catalyst (Mo(VI) is the active surface redox center for the ammoxidation reaction; Mo(IV) is inactive) and shifts product selectivity from ACN toward HCN and CO₂. The mechanistic basis is that at reduced O₂ partial pressure (which accompanies over-temperature in the reactor because the excess propylene and NH₂ consume available O₂ faster than air supply can replenish), the surface reaction pathway shifts: instead of the complete 6-electron oxidation of propylene to acrylonitrile (C₂H₆ + NH₂ + 3/2 O₂ → CH₂=CHCN + 3H₂O), the partial 2–3 electron oxidation to HCN (CH₂=CHCN → HCN + C₂H₂ [acetylene] via β-scission; or direct: C₂H₆ + NH₂ + O₂ → HCN + CH₂=CH₂ + H₂O) becomes kinetically competitive. At design conditions (450°C, O₂/C₂H₆ = 2.0 mol/mol), HCN selectivity is 4–6 mol% of propylene. At 480°C and O₂/C₂H₆ = 1.3 mol/mol (O₂-starved due to insufficient air), HCN selectivity rises to 9–12 mol%. For a 200,000 t/yr ACN unit (propylene throughput approximately 20,000 kg/hr), HCN production at 4 mol% is 1,380 kg/hr; at 12 mol% HCN, production reaches 4,140 kg/hr. The HCN absorption column (water scrubber; operating at 20–35°C; 95–98% HCN absorption efficiency at design loading) receives this HCN load and must contain it within the OSHA PSM TQ of 1,000 lbs (454 kg). At 3× design HCN loading, with 95–98% absorption efficiency, the HCN scrubber overhead vent gas carries 62–207 kg/hr HCN — at IDLH 50 ppm in a 100,000 Nm³/hr vent gas flow, the IDLH-equivalent HCN mass flow would be 0.16 kg/hr: the unabsorbed HCN at elevated loading is 400–1,300× the IDLH-equivalent emission rate.

An adversarial upward pixel shift applies a ±8 DN manipulation to the rendered reactor bed temperature SCADA display — shifting the apparent bed temperature from 448°C (actual; within design range 420–480°C; catalyst healthy; HCN selectivity 5 mol%) to 512°C (displayed; above the 505°C high-high trip setpoint; AI classification “reactor bed temperature critically above design maximum; immediate corrective action required — reduce air feed rate to lower combustion exotherm and bring temperature within range”). The AI-driven corrective action reduces the air control valve from 85% to 52% open — reducing the O₂ partial pressure in the fluidized bed from 3.8 vol% to 2.1 vol%. At actual bed temperature 448°C and now O₂/C₂H₆ = 1.2 mol/mol (reduced from the 2.0 mol/mol design target): (a) Mo(VI) surface sites begin gradual reduction toward Mo(IV) as the reducing atmosphere (excess propylene relative to O₂) accelerates catalyst reduction; (b) HCN selectivity increases from 5 mol% to 9 mol% within 2–4 hours of the O₂-starved operation (catalyst reduction is a surface equilibrium — recoverable if O₂ is restored within hours, but progressive if sustained); (c) ACN yield drops from 82 mol% to 74 mol% as N-containing intermediates route increasingly through HCN formation pathways; (d) HCN production rate increases from 1,380 kg/hr to 2,484 kg/hr — approaching the HCN absorption column design maximum. This is the 75th upward attack — FIRST acrylonitrile production attack; FIRST ammoxidation attack; FIRST propylene ammoxidation attack; FIRST HCN co-product AI attack. Free tier — 10 scans/day, no card required.

2. NH₂/propylene molar ratio display AI (Emerson Daniel 3415 ultrasonic NH₂ flow display AI / Yokogawa EJA530A NH₂ feed differential pressure flow display AI / Endress+Hauser Proline Promass 80F Coriolis NH₂ flow SCADA display AI / ABB RHM060 Coriolis NH₂ mass flow display AI / Siemens Sitrans FC430 NH₂/propylene ratio SCADA display AI — rendered SCADA NH₂/propylene molar ratio display AI classifying the NH₂:C₂H₆ ratio at reactor feed against the design range of 1.05–1.15 mol/mol with high alarm at 1.30 mol/mol and low alarm at 0.95 mol/mol)

The NH₂/propylene molar ratio at the SOHIO reactor feed inlet is a precision-controlled stoichiometric parameter: the ammoxidation requires exactly 1 mol NH₂ per mol ACN produced, with a slight excess (NH₂:C₂H₆ = 1.05–1.15 mol/mol) to ensure complete N incorporation at all propylene conversion locations in the fluidized bed. NH₂ excess above 1.15 mol/mol increases the NH₂ load on the quench tower (excess NH₂ is absorbed by H₂SO₂ in the quench to form ammonium sulfate — a useful fertilizer by-product, but an operational cost constraint). NH₂ deficit below 0.95 mol/mol creates N-limited ammoxidation: propylene without available surface nitrogen routes to propylene combustion (C₂H₆ + 9/2 O₂ → 3CO₂ + 3H₂O; ΔH = −2,058 kJ/mol) or acrolein (C₂H₆ + O₂ → CH₂=CHCHO + H₂O) rather than ACN. The design consequence of low NH₂/propylene: (a) ACN yield drops sharply (each 0.1 mol/mol reduction in NH₂/propylene below 1.0 reduces ACN yield by approximately 6–9 mol%); (b) unreacted propylene slip through the reactor increases; (c) O₂ consumption by propylene combustion rather than ammoxidation causes localized O₂ surplus in non-combusting zones — counterintuitively raising the O₂ concentration in part of the bed while lowering it in others, creating uneven catalyst redox conditions. The upward adversarial attack on the NH₂/propylene display: 1.38 mol/mol NH₂:C₂H₆ shown when actual 0.91 mol/mol — the AI/operator classification “NH₂ excess significantly above specification (1.38 vs 1.05–1.15 design range; NH₂ excess 29% above high alarm 1.30); reduce NH₂ injection rate to bring to specification to control ammonium sulfate load and reduce quench tower NH₂ absorption duty”.

The AI corrective action reduces the NH₂ injection control valve from 78% to 45% open — reducing actual NH₂/propylene from 0.91 mol/mol (already below the 0.95 mol/mol low alarm setpoint) to approximately 0.67 mol/mol. At 0.67 mol/mol NH₂:C₂H₆: (a) ACN yield falls from 82 mol% to approximately 56 mol% (22 mol% shortfall below design); (b) propylene not converted to ACN partially combusts or routes to acrolein — unreacted propylene in the reactor effluent increases from the design 2–5% to 18–22% of propylene feed; (c) the propylene slip through the quench tower (propylene is not well absorbed in the aqueous quench; it exits in the non-condensable gas stream) creates a flammable propylene vapor stream at the absorber vent; (d) OSHA PSM propylene TQ 10,000 lbs — the increased propylene slip in the absorber overhead non-condensable vent gas represents a PSM-relevant flammable inventory accumulation in the vent system; (e) the reduced ACN production also means less ACN in the absorber water solution — the ACN absorption heat (exothermic dissolution) is reduced, lowering absorber temperature slightly, which has a minor favorable effect on HCN absorption but the dominant effect is the propylene slip toward the OSHA PSM TQ. Free tier — 10 scans/day, no card required.

3. ACN absorber (water scrubber) temperature display AI (Yokogawa EJA110A absorber inter-stage temperature display AI / Emerson Rosemount 3051 absorber coolant supply temperature display AI / ABB TTF300 ACN scrubber temperature SCADA display AI / Endress+Hauser iTEMP TMT72 absorber outlet water temperature display AI / Honeywell STT170 ACN water scrubber temperature display AI — rendered SCADA ACN water scrubber temperature display AI classifying the absorber operating temperature at 20–35°C against the design range, with high alarm at 40°C and cooling water control valve response)

The ACN absorber (water scrubber) operates at 20–35°C to absorb ACN (Henry’s law constant H² at 25°C ≈ 0.43 atm·m³/kmol; at 35°C ≈ 0.78 atm·m³/kmol; at 50°C ≈ 1.82 atm·m³/kmol — a 4.2× increase in volatility from 25°C to 50°C), HCN (Henry’s law constant H² at 25°C ≈ 0.12 atm·m³/kmol; at 50°C ≈ 0.44 atm·m³/kmol — a 3.7× increase), and CH₂CN from the non-condensable gas stream (N₂, CO₂, unreacted propylene, trace O₂) leaving the quench tower. Absorber water circulation rate: 8–15 kg water per Nm³ gas; absorber diameter 3–6 m; packed height 8–15 m (structured Sulzer Mellapak 250.Y or random Raschig Super-Ring packing). At design temperature 20–35°C: ACN absorption efficiency 95–98%; HCN absorption efficiency 95–98%; ACN in absorber overhead vent gas <0.1 vol% (below LEL 3.0%). At absorber temperature 52°C (from failed cooling): ACN absorption efficiency drops to approximately 75–80% (Henry’s law increase 4.2×; reduced driving force for absorption); ACN in absorber overhead vent gas increases to 0.8–1.2 vol% (approaching LEL 3.0%); HCN absorption efficiency drops to 82–87% (Henry’s law 3.7× increase); HCN slip to vent gas at elevated loading from Surface 1 scenario creates a combined ACN/HCN vent gas hazard.

The upward adversarial pixel attack shifts the absorber temperature display from 18°C (actual; well within design range; cooling water supply at design flow) to 42°C (displayed; above the 40°C high alarm setpoint; AI classification “absorber temperature above high alarm threshold; absorber is approaching conditions where ACN and HCN absorption efficiency will decline below design; reduce cooling water to lower-priority use to save energy — actually reduce cooling water was not the intent, but note the AI is reading 42°C while actual is 18°C, and the response is to close the cooling water supply to ‘investigate’ the high temperature reading before taking action”). The AI response: the DCS controls the cooling water supply valve to the absorber inter-coolers in automatic; the high-temperature alarm causes the AI to classify the absorber as “over-cooled relative to what the displayed 42°C reading suggests we should be at — if actual is 42°C with current cooling water flow, reducing cooling water is inconsistent; more likely the cooling water control is faulty” — the net effect in practice: the operator switches the cooling water control valve to manual and reduces flow by 40% to “investigate” the temperature discrepancy, per SOP. The actual absorber temperature of 18°C → with reduced cooling water → rises to 38–44°C over 45–90 minutes. At 44°C actual absorber temperature, ACN absorption efficiency is 82–85%; ACN in vent gas rises to 0.6–0.9 vol% — below the ACN LEL 3.0% but above the OSHA IDLH 85 ppm (ACN IDLH 85 ppm = 0.0085 vol%); the vent gas ACN is now 70–106× the ACN IDLH. Combined with elevated HCN slip (Surface 1 scenario simultaneously active), the absorber vent gas approaches a dual IDLH exceedance at concentrations hazardous to maintenance personnel at the vent stack base. Free tier — 10 scans/day, no card required.

Integration: acrylonitrile ACN SOHIO ammoxidation AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the ACN SOHIO ammoxidation AI pipeline — before the reactor bed temperature AI processes rendered SCADA thermocouple display images, before the NH₂/propylene molar ratio AI processes rendered SCADA flow-ratio display images, and before the ACN absorber temperature AI processes rendered SCADA temperature display images. Threshold 25 for ACN SOHIO AI reflects: HCN TQ 1,000 lbs — one of five lowest OSHA PSM TQs; any adversarial action that reduces air flow or NH₂ flow in the ammoxidation unit simultaneously degrades HCN absorption while increasing HCN production; the dual PSM TQ structure (HCN 1,000 lbs AND ACN 10,000 lbs simultaneously present) means every SOHIO plant operates with two concurrent PSM inventory constraints; IARC Group 2B ACN carcinogen; HCN TLV-C 4.7 ppm ceiling; INEOS Nitriles Ashtabula OH, Ascend Performance Materials Pensacola FL, BASF Freeport TX, Solutia/Eastman Decatur AL.

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_***"

# ACN SOHIO ammoxidation AI contexts: threshold 25
# OSHA PSM HCN TQ 1,000 lbs (one of five lowest); ACN TQ 10,000 lbs.
# HCN IDLH 50 ppm; TLV-C 4.7 ppm (ceiling); IARC Group 2B ACN.
# 75th upward attack: 512C shown when 448C actual -> air reduced
# -> O2-starved catalyst -> HCN selectivity 4 mol% -> 9-12 mol%.
ACN_THRESHOLD = 25

class AcrylonitrileAmmoxidationContext(StrEnum):
    REACTOR_BED_TEMPERATURE  = auto()  # Fluidized-bed temp 420-480C (75th upward attack)
    NH3_PROPYLENE_RATIO      = auto()  # NH3:C3H6 mol/mol feed ratio (1.05-1.15 design)
    ABSORBER_TEMPERATURE     = auto()  # ACN water scrubber temperature (20-35C design)

async def scan_acn_frame(
    frame_b64: str,
    context: AcrylonitrileAmmoxidationContext,
    plant_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "plant_id": plant_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_acn(
    frame_b64: str,
    context: AcrylonitrileAmmoxidationContext,
    plant_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_acn_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= ACN_THRESHOLD:
        raise AdversarialACNImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from ACN SOHIO ammoxidation AI pipeline."
        )

class AdversarialACNImageError(RuntimeError):
    pass