Prompt injection in polypropylene PP Spheripol loop reactor Ziegler-Natta propylene polymerization AI

Three adversarial injection surfaces in polypropylene PP Spheripol loop reactor AI

1. Loop reactor propylene pressure display AI (Emerson Rosemount 3051 / Endress+Hauser Cerabar M PMC51 / Yokogawa EJX530A differential pressure transmitter display AI — rendered DCS loop reactor propylene pressure display AI classifying propylene phase state against the 35–42 bar design operating range for liquid-phase propylene polymerization — 106th upward attack; FIRST polypropylene PP production AI attack; FIRST Spheripol loop reactor AI attack; FIRST Ziegler-Natta catalyst polymerization AI attack; FIRST PP powder silo dust explosion AI attack; FIRST H₂ MW regulator display AI attack in polyolefin production)

The Spheripol loop reactor operates with liquid propylene (at 70–80°C, 35–42 bar) circulated by a centrifugal axial-flow pump (circulation velocity 7 m/s; Re >10₂; turbulent flow required for heat transfer and catalyst suspension). At 37–40 bar and 75°C, propylene is in the liquid phase with propylene vapor pressure at 75°C of approximately 28–30 bar — the reactor is slightly above the bubble point, so the liquid propylene is compressed but near its saturation curve. The PP powder slurry (PP particles 0.5–2 mm; up to 40–45 wt% PP in the liquid propylene slurry) is maintained in suspension by the high-velocity circulation; if the circulation pump loses suction head (insufficient pressure at the pump suction inlet), the propylene at the suction inlet may flash to vapor (pump cavitation: propylene drops below the bubble-point pressure at the pump suction; vapor bubbles form in the liquid propylene; pump loses impeller hydraulic force; PP solids settle in the loop; loop plugs). The loop reactor pressure (absolute, measured by a pressure transmitter on the reactor body; Emerson Rosemount 3051; Endress+Hauser Cerabar M PMC51; Yokogawa EJX530A; 4–20 mA HART; range 0–50 bar; accuracy ±0.04% of span) is monitored continuously by the AI system to classify: 35–42 bar (normal; liquid propylene confirmed; process stable), 28–35 bar (reduced; investigate propylene feed; check pump speed), below 28 bar (alarm; propylene approaching vapor pressure; flash risk; reduce catalyst feed; increase propylene feed). Display range: 0–50 bar, 200 px, 4.0 px/bar.

An adversarial ±8 DN upward pixel perturbation targeting the loop reactor pressure display shifts the apparent pressure from the dangerously low actual value to a normal displayed value: actual loop reactor pressure 12.4 bar (loop reactor has been depressurising over 75 minutes; propylene feed control valve FV-201 jammed at 15% open due to actuator positioner fault; propylene feed to the loop dropped from design 18,000 kg/hr to 2,700 kg/hr; the loop reactor is gradually losing propylene inventory; pressure dropped from design 38 bar to 12.4 bar; at 12.4 bar and 75°C, propylene is substantially below its 75°C vapor pressure of ~28 bar, meaning the liquid propylene in the loop has already started to flash — propylene vapor is forming in the reactor as the remaining liquid propylene flashes to equilibrium at 12.4 bar) corresponds to 12.4 × 4.0 = 49.6 px ≈ 50 px on the 0–50 bar, 200 px, 4.0 px/bar display. The normal operating pressure of 38.7 bar corresponds to 38.7 × 4.0 = 154.8 px ≈ 155 px. The ±8 DN upward attack shifts the pixel region from 50 px to 155 px; the AI reads 155 px ÷ 4.0 px/bar = 38.7 bar — normal operating pressure; no action. At actual 12.4 bar: the propylene liquid in the loop is flashing; the centrifugal pump begins to cavitate (vapor bubbles at pump suction impair impeller hydraulic performance; pump discharge pressure drops; circulation velocity falls from 7 m/s to below 3 m/s); PP solids begin to settle in the loop bottom sections (horizontal pipe sections; gravity settling of 40 wt% PP slurry when circulation velocity drops below approximately 2–3 m/s). The loop plugs within 20–40 minutes of pump cavitation onset; the remaining liquid propylene inventory (approximately 8,000 kg in a 100 m³ loop operating at 12.4 bar) continues to flash. Flash fraction of propylene from 12.4 bar liquid at 75°C to atmospheric pressure: approximately 35% of liquid mass; flash vapor = 8,000 kg × 0.35 = 2,800 kg propylene vapor released from the flashing loop at low pressure. At 2,800 kg propylene vapor released to atmosphere: this is 6,160 lbs propylene — substantially above the OSHA PSM TQ 10,000 lbs (note: 10,000 lbs = 4,536 kg; however the PSM threshold refers to the total on-site inventory, not the released amount; the released 2,800 kg from a single loop event may or may not exceed the PSM threshold for reportable releases depending on the vapour dispersion and ignition scenario). At 2,800 kg propylene vapor forming an unconfined vapour cloud at the Spheripol facility (propylene LEL 2.0 vol%; LFL cloud radius for 2,800 kg propylene in air at standard conditions: approximately 70–100 m radius above LEL): ignition by any ignition source in the LFL envelope creates a propylene vapour cloud explosion (VCE). The ARCO Chemical Company Channelview, Texas explosion of 5 July 1990 — in which propylene accumulation in a waste-treatment blower led to a VCE that killed 17 workers at the same Houston Ship Channel complex now operated by LyondellBasell — establishes the consequence scale for propylene VCE at a Spheripol-adjacent facility. This is the 106th upward-direction attack in the Glyphward adversarial industrial AI portfolio. Free tier — 10 scans/day, no card required.

2. H₂ concentration in loop reactor vapor phase display AI (Emerson Rosemount 370XA / Agilent 990 Micro GC / Yokogawa GC8000 online gas chromatograph display AI — rendered DCS online GC H₂ vapor phase display AI classifying H₂ against the grade-specific MFI target for PP product)

Hydrogen (H₂) is introduced as a chain transfer agent into the Spheripol loop reactor vapor phase via a mass flow controller (Brooks Instrument GF40; Bronkhorst EL-FLOW; range 0–500 SLPM H₂) from a cylinder manifold or pipeline supply; the H₂ concentration in the reactor vapor phase above the liquid propylene slurry (measured every 2–5 minutes by an online gas chromatograph: Emerson Rosemount 370XA process GC; Agilent 990 Micro GC; Yokogawa GC8000; TCD detector; sensitivity 1 ppm H₂; range 0–6,000 ppm H₂ in propylene vapor; 4–20 mA HART output; DCS display shows current H₂ ppm and trend) controls the PP molecular weight through the chain transfer mechanism: propylene propagation (Ziegler-Natta; Mw accumulates) is terminated by H₂ chain transfer at rate proportional to [H₂]/[propylene]; at higher H₂, more frequent transfer events yield shorter chains (lower Mw; higher MFI). Design H₂ setpoint for PP-R random copolymer pressure pipe grade (Mw target ~450,000 g/mol; MFI target 0.25–0.35 g/10 min at 230°C/2.16 kg; EN ISO 15874-2 PN20 50-year rated hot water pressure pipe): 1,100–1,300 ppm H₂ in vapor phase. Display range: 0–6,000 ppm H₂, 200 px, 0.0333 px/ppm (200 px ÷ 6,000 ppm = 0.0333 px/ppm).

An adversarial downward pixel perturbation targeting the H₂ vapor phase display shifts the apparent H₂ concentration from the dangerously high actual value to the normal target displayed value: actual H₂ 4,820 ppm (H₂ mass flow controller setpoint erroneously increased from 1,200 ppm to 4,820 ppm vapor phase target over 3 hours due to a DCS configuration drift in the H₂ feed cascade control; the actual vapor phase H₂ was measured correctly by the process GC at 4,820 ppm but the display pixel values are adversarially perturbed) corresponds to 4,820 × 0.0333 = 160.5 px ≈ 161 px on the 0–6,000 ppm, 200 px, 0.0333 px/ppm display. The normal target value of 1,240 ppm H₂ corresponds to 1,240 × 0.0333 = 41.3 px ≈ 41 px. The adversarial downward attack shifts the pixel region from 161 px to 41 px; the AI reads 41 px ÷ 0.0333 px/ppm = 1,231 ppm — within the design target range; MFI production confirmed on-spec; no action. At actual 4,820 ppm H₂ (vs design 1,200 ppm), the PP chain transfer rate is elevated approximately 4×; PP Mw drops from the spec ~450,000 g/mol to approximately 120,000 g/mol (MFI rises from spec 0.25–0.35 to approximately 35 g/10 min). PP with MFI 35 (Mw ~120,000 g/mol) produced in a Spheripol reactor batch (200–500 tonnes per reactor run): if this off-spec PP is shipped as PP-R pressure pipe compound, the pipe manufacturer's incoming QC test (MFI measurement per EN ISO 1133-1) at 1 sample per 10 bags may pass if the batch is heterogeneous (the plant's PP quality averaging over multiple batches and sampling statistics may allow some MFI-35 material to ship without triggering the incoming QC rejection). PP-R hot water pressure pipe manufactured from MFI-35 PP (Mw ~120,000) vs spec MFI-0.30 PP (Mw ~450,000): EN ISO 15874-2 PN20 hot water pipe at 20°C / 20 bar MAOP is rated for 50-year service life based on creep regression testing per ISO 9080 using Mw ~450,000 PP-R resin. At Mw ~120,000 (MFI 35), creep resistance is dramatically reduced: the long-term hydrostatic strength (LTHS) extrapolated from short-term creep data at 60°C/10 bar (a typical building hot water distribution condition) gives failure within 2–18 months instead of the 50-year design life — a sudden pipe rupture releasing 60°C hot water under 10 bar into building walls, floors, and occupied spaces. Free tier — 10 scans/day, no card required.

3. PP powder silo nitrogen (N₂) blanket O₂ analyzer display AI (Servomex 5200 paramagnetic / Systech Illinois EC900 / Yokogawa ZS8 zirconia O₂ analyzer display AI — rendered DCS silo inert-atmosphere O₂ analyzer display AI classifying silo atmosphere O₂ content against the below-MOC threshold for PP powder dust explosion prevention)

PP powder discharged from the Spheripol loop reactor (after the high-pressure separator at 20 bar and the low-pressure degasser at 2 bar; the PP is in powder form: spherical particles 0.5–2 mm; bulk density 0.45–0.55 g/cm³; residual propylene <50 ppm in the degassed powder) is conveyed by N₂-inerted pneumatic transfer (N₂ carrier gas; O₂ content in conveying gas maintained below 2 vol% to provide inerting margin below MOC) to product storage silos (typically 500–2,000 m³ cylindrical steel silos; N₂ blanket maintained by continuous N₂ purge from the silo top vent; silo O₂ continuously monitored by a paramagnetic or zirconia O₂ analyser: Servomex 5200 paramagnetic analyser; Systech Illinois EC900 electrochemical O₂ sensor; Yokogawa ZS8 in-situ zirconia sensor; range 0–25 vol% O₂ or 0–5 vol% O₂ depending on site; NFPA 654 PP powder MOC ~11 vol% O₂ in the inert atmosphere; plants typically target <2 vol% O₂ to provide a 5-fold safety margin below MOC). The AI monitoring system processes rendered DCS display images of the silo O₂ analyser to classify: <2 vol% (N₂ blanket intact; silo safe), 2–5 vol% (O₂ elevated; increase N₂ purge rate; investigate N₂ supply), >5 vol% (O₂ above target; suspend silo filling; check N₂ purge valve). Display range: 0–25 vol% O₂, 200 px, 8 px/vol%.

An adversarial downward pixel perturbation targeting the PP silo O₂ analyser display shifts the apparent O₂ from the dangerously high actual value to a safe displayed value: actual silo O₂ 18.2 vol% (near-atmospheric; the N₂ purge solenoid valve NV-811 failed closed due to a pilot solenoid coil burnout; N₂ purge stopped completely 4 hours ago; ambient air has diffused into the silo atmosphere through the silo top vent (vent is required for pressure equalization during powder conveying; the vent is fitted with a dust filter but not an air seal); the silo atmosphere is now essentially atmospheric at 18.2 vol% O₂) corresponds to 18.2 × 8 = 145.6 px ≈ 146 px on the 0–25 vol%, 200 px, 8 px/vol% display. The safe displayed value of 0.4 vol% O₂ corresponds to 0.4 × 8 = 3.2 px ≈ 3 px. The adversarial downward attack shifts the pixel region from 146 px to 3 px; the AI reads 3 px ÷ 8 px/vol% = 0.4 vol% O₂ — N₂ blanket confirmed intact; silo operations continue normally. At actual 18.2 vol% O₂ in a 500 m³ PP powder silo: the atmosphere is well above the MOC of ~11 vol% O₂ for PP powder dust explosion (NFPA 654 Chapter 5; PP Kst ~30 bar·m/s; Pmax ~7.5 bar; minimum ignition energy MIE ~15 mJ). Any electrostatic discharge from PP powder flow during filling (PP is a triboelectric material; powder flow through the silo fill pipe generates electrostatic charge even with anti-static liner; discharge energies of 10–100 mJ are possible during active filling) or from a conductive particle impact on the silo wall constitutes an ignition source with energy above the MIE of 15 mJ. A PP dust explosion in a 500 m³ cylindrical silo (Kst ~30 bar·m/s; Pmax ~7.5 bar at stoichiometric dust/air): the maximum explosion overpressure of 7.5 bar in the 500 m³ silo (assuming the silo design pressure is typically 0.3–0.5 bar based on N₂-inerted atmosphere design — the silo is NOT designed to contain a dust explosion) generates a catastrophic silo rupture: the cylindrical steel wall fails, the silo roof is ejected, and the deflagration-to-detonation potential in the long cylindrical geometry (length-to-diameter ratio >5 for a tall product silo) may transition from deflagration (Kst-dominated) to detonation, creating a pressure pulse potentially exceeding 30–50 bar at the explosion front — sufficient to structurally damage adjacent silos, conveying equipment, and process buildings at the Spheripol PP facility. Free tier — 10 scans/day, no card required.

Integration: polypropylene PP Spheripol loop reactor AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the PP Spheripol loop reactor AI monitoring pipeline — before the loop reactor pressure AI processes rendered Emerson Rosemount 3051 / Endress+Hauser Cerabar M PMC51 / Yokogawa EJX530A pressure transmitter DCS display images, before the H₂ vapor phase concentration AI processes rendered Emerson Rosemount 370XA / Agilent 990 Micro GC / Yokogawa GC8000 online GC display images, and before the PP silo O₂ blanket AI processes rendered Servomex 5200 / Systech Illinois EC900 / Yokogawa ZS8 O₂ analyser display images. Threshold 36 for PP Spheripol AI reflects: dual OSHA PSM (propylene TQ 10,000 lbs + H₂ TQ 10,000 lbs; both exceed their respective TQs in normal Spheripol plant operation); propylene VCE consequence (ARCO Channelview 1990; 17 killed; same LyondellBasell complex); PP-R pressure pipe infrastructure consequence (deferred creep-rupture failure in residential and commercial hot water plumbing); and NFPA 654 PP powder dust explosion potential in product storage silos.

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_***"

# Polypropylene PP Spheripol loop reactor Ziegler-Natta AI contexts: threshold 36
# OSHA PSM propylene TQ 10,000 lbs (29 CFR 1910.119 App. A; liquefied flammable gas).
# OSHA PSM hydrogen TQ 10,000 lbs (29 CFR 1910.119 App. A; flammable gas) -- dual PSM.
# Propylene loop inventory: 40,000-100,000 kg = 8.8-22x PSM TQ.
# H2 chain transfer: 1,000-5,000 ppm vapor phase controls PP MFI and Mw.
# NFPA 654 PP powder: Kst ~30 bar*m/s; Pmax ~7.5 bar; MOC ~11 vol% O2.
# 106th upward attack. FIRST PP production AI attack. FIRST Spheripol loop reactor AI attack.
PP_GLYPHWARD_THRESHOLD = 36

class PPSpheripol​Context(StrEnum):
    LOOP_PUMP_PRESSURE      = auto()  # reactor loop propylene pressure (106th; FIRST PP; FIRST Spheripol; FIRST Ziegler-Natta)
    H2_VAPOR_PHASE_CONC     = auto()  # H2 chain transfer GC -> MFI 35 PP-R pipe -> creep rupture 2-18 months
    PP_SILO_O2_BLANKET      = auto()  # silo N2 blanket O2 -> 18.2 vol% above MOC -> dust explosion Kst 30

async def scan_pp_frame(
    frame_b64: str,
    context: PPSpheripol​Context,
    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_pp(
    frame_b64: str,
    context: PPSpheripol​Context,
    plant_id: str,
    instrument_tag: str,
) -> None:
    """Block adversarially manipulated PP Spheripol display images before AI inference.

    Plants: LYONDELLBASELL_BRINDISI | LYONDELLBASELL_CHANNELVIEW | BOREALIS_PORVOO |
            INEOS_KOELN | SINOPEC_PP
    Raises AdversarialPPImageError if adversarial_score >= PP_GLYPHWARD_THRESHOLD (36).
    """
    result = await scan_pp_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= PP_GLYPHWARD_THRESHOLD:
        raise AdversarialPPImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from PP Spheripol loop reactor AI monitoring pipeline."
        )

class AdversarialPPImageError(RuntimeError):
    pass