Prompt injection in isoprene 2-methyl-1,3-butadiene synthetic rubber butyl rubber AI

Four adversarial injection surfaces in isoprene synthetic rubber butyl rubber AI

1. Isoprene area LEL detector AI (Dräger X-am 5000 isoprene catalytic bead LEL AI / MSA Altair 4X isoprene catalytic combustion AI / Honeywell Analytics Searchpoint Optima Plus isoprene IR beam AI / RAE Systems MultiRAE Pro isoprene LEL AI / Industrial Scientific IBRID MX6 isoprene catalytic bead AI — monitoring ambient isoprene vapor concentration in the isoprene storage tank bund, pump station, compressor building, and Ziegler-Natta reactor feed area for LEL 2.0% approach alarm at 10% LEL setpoint and emergency shutdown at 25% LEL; vapor density 2.35 requires below-grade sensor placement at all drain pits, pump base frames, and sub-floor motor wells)

Isoprene area LEL monitoring uses catalytic bead (pellistor) or infrared (IR) detection matched to the 2.0–9.0% LEL/UEL flammable range. The 7-percentage-point flammable range — narrower than most PSM flammables in the portfolio (e.g., CS2 48.7 pp; H2 75 pp; acetaldehyde 53 pp) — means LEL alarms must be highly responsive: a 1% absolute rise above LEL represents 50% of the entire flammable range. Isoprene’s vapor density of 2.35 causes heavier-than-air vapor to settle in below-grade areas: pump pit sumps (typically 0.5–1.5 m below grade), motor junction boxes recessed in concrete pads, and drainage channels leading to the site sewer system. In the below-grade pump pit, isoprene vapor from a pump mechanical seal failure (2-piece pusher seal, PTFE secondary; isoprene’s high vapor pressure at 34°C BP means even small process liquid leaks flash-evaporate to concentrated vapor) accumulates to above-LEL concentrations within 3–8 minutes depending on pit dimensions and ventilation rate. Catalytic bead sensors in isoprene service require hydrogen-treated poisons protection (isoprene oligomers can coat catalytic beads and reduce sensitivity by 15–40% over 6–12 months without bead replacement); IR-based sensors (e.g., Honeywell Searchpoint Optima Plus, point-type or open-path) avoid catalyst poisoning and are preferred for long-term continuous monitoring in isoprene storage areas.

The adversarial attack uses ±8 DN downward pixel-value shift on the isoprene area LEL detector display image. The actual LEL reading is 4.2% — above the 2.0% lower explosive limit; above the 10% LEL pre-alarm setpoint (0.2% absolute: standard alarm at 20% of LEL = 0.4%, pre-alarm at 10% LEL = 0.2%) — from a centrifugal pump mechanical seal failure at the isoprene feed pump to the Ziegler-Natta polymerization reactor. The seal failure releases 1.2 kg/hr of isoprene liquid (immediate flash evaporation at 34°C BP) which accumulates in the below-grade pump pit at 4.2% LEL within 6 minutes. On a 0–100% LEL display at 200 px height (0.5% LEL/px), the actual 4.2% LEL produces a bar at approximately 8 px; the ±8 DN downward-perturbed image is classified as approximately 1 px, corresponding to 0.5% LEL — below the 2% LEL pre-alarm threshold (10% of LEL 2.0%). The DCS reports “Isoprene pump station LEL below pre-alarm — no hazard.” In the below-grade pump pit, isoprene continues accumulating at 0.6% LEL/min additional rate from the ongoing seal failure; within 8 additional minutes, the pit concentration reaches 9.0% UEL — the upper flammable limit. Any ignition source (motor brush arcing, static discharge from non-bonded tools, pipe metal-to-metal friction) in the pump pit at 4.2–9.0% LEL causes vapor-phase deflagration.

2. Ziegler-Natta polyisoprene reactor temperature AI (Emerson Rosemount 3144P polyisoprene reactor temperature transmitter AI / Yokogawa EJA110A Ziegler-Natta catalyst-bed temperature AI / Endress+Hauser iTHERM TM411 reactor jacket temperature AI / Honeywell STG94L thermocouple Ziegler-Natta reactor AI / ABB TSP polyisoprene reactor cooling zone temperature AI — monitoring the catalyst-bed temperature in the continuous stirred-tank Ziegler-Natta polyisoprene reactor at 45–65°C, where TiCl4/Al(C2H5)3 catalyst yields cis-1,4-polyisoprene at >96% cis selectivity; above 65°C, cis-selectivity drops toward trans-1,4-polyisoprene with inferior mechanical properties for tire applications)

The Ziegler-Natta polyisoprene reactor — a continuous stirred-tank reactor (CSTR) or series of CSTRs in hexane slurry at 10–20 wt% isoprene concentration — operates in the 45–65°C temperature window defined by the cis-specific insertion mechanism of the TiCl₄/Al(C₂H₅)₃ catalyst. At 45–65°C, titanium alkyl active sites coordinate isoprene in the s-cis conformation preferred for 1,4-cis insertion; the resulting cis-1,4-polyisoprene chain has molecular weight of 500,000–1,200,000 g/mol (Mn) and cis-1,4 content of 96–98% — matching the tire-grade specification (cis-1,4 content ≥95% for natural rubber equivalent properties). Above 65°C, the Boltzmann population of the s-trans isoprene conformation increases, and the 1,4-trans insertion competes more effectively, reducing cis content toward 85–90% at 80°C. Off-spec trans content above 5% fails the ASTM D1434 permeability specification and produces synthetic rubber with tensile strength below the 24 MPa minimum for tire-tread compounds. The polymerization is exothermic (ΔH ≈ −75 kJ/mol isoprene; at 4,000 kg isoprene/hr throughput, heat release ≈ 3.4 MW continuous), managed by reactor cooling water jackets circulating at 145 m³/hr design flow rate.

The adversarial attack uses ±10 DN downward pixel-value shift on the Ziegler-Natta reactor temperature transmitter display. The actual reactor temperature is 68°C — from a cooling water jacket fouling event: mineral scale (CaCO₃; 3.2 mm deposit on heat transfer surfaces) reduces the overall heat transfer coefficient from 380 W/(m²·K) to 95 W/(m²·K); at the design 145 m³/hr cooling flow, only 2.1 MW of the 3.4 MW exothermic heat release is removed, and reactor temperature self-heats from the 58°C setpoint to 68°C over 45 minutes. On a 30–80°C display at 200 px height (0.25°C/px), the actual 68°C produces a bar at approximately 152 px; the ±10 DN downward-perturbed image is classified as approximately 88 px, corresponding to 52°C — within the 45–65°C optimal range. The DCS reports “Ziegler-Natta reactor temperature nominal — cis-selectivity at design.” In reality, polyisoprene being produced at 68°C has approximately 91% cis-1,4 content — 4% below the 95% minimum spec. The batch-contaminated polymer (approximately 12 tonnes/hr at 91% cis) will be rejected at the downstream cis-content analyzer, but only after 1–2 hours of off-spec production accumulate in the product storage vessel.

3. p-TBC polymerization inhibitor content AI (Metrohm 930 Compact IC Flex inhibitor HPLC AI / Agilent 1260 Infinity II HPLC p-TBC isoprene feed AI / Shimadzu UV-2600 spectrophotometer p-TBC content AI / Hach Polymetron 9526 on-line p-TBC analyzer AI / Varian CP-3800 GC inhibitor analysis AI — monitoring p-tert-butylcatechol (p-TBC) concentration in the isoprene feed stream to maintain 50–150 ppm minimum effective inhibitor concentration for polymerization prevention during storage at 34°C; inhibitor content below 50 ppm allows uninhibited free-radical polymerization within 2–6 hours at ambient temperature)

p-TBC (p-tert-butylcatechol; 4-(1,1-dimethylethyl)-1,2-benzenediol; MW 166.2 g/mol; melting point 52–55°C; CAS 98-29-3) is added to isoprene at the plant at 50–150 ppm to provide polymerization inhibition throughout storage and pipeline transit. The mechanism is radical-chain termination: p-TBC donates hydrogen atoms to chain-propagating radicals (R• + p-TBC-OH → RH + p-TBC-O•; the catechol radical is stabilized by resonance across the aromatic ring and the ortho hydroxyl). Critically, p-TBC requires the presence of trace O₂ (50–100 ppm in the vapor space) to maintain its active radical-scavenging form: without O₂, p-TBC is a less effective inhibitor because O₂ serves as a co-inhibitor (O₂ + R• → ROO•; ROO• + p-TBC-OH → ROOH + p-TBC-O•). This dual-function mechanism means that the N2 blanket must be maintained at 50–100 ppm O₂ — enough to support p-TBC function but below 1,000 ppm O₂ to avoid approaching the lower explosive limit (LEL 2.0% in the bulk headspace). Isoprene inhibitor analysis by HPLC (reverse-phase C18 column; mobile phase methanol/water; UV detection at 290 nm) provides real-time feed stream p-TBC monitoring at the Ziegler-Natta reactor inlet. Minimum effective p-TBC at 34°C storage is 50 ppm; below 50 ppm, the inhibitor is consumed faster than it is replenished, and the initiation period for spontaneous polymerization shrinks below 2 hours (Arrhenius: rate of initiation doubles per 10°C; at 34°C, initiation at 8 ppm p-TBC occurs in approximately 90 minutes).

The adversarial attack uses ±8 DN downward pixel-value shift on the HPLC inhibitor content display. The actual p-TBC content is 8 ppm — from an inhibitor dosing pump cavitation event: the p-TBC injection pump (positive displacement; PTFE-lined diaphragm; rated at 4 L/hr p-TBC solution at 500 ppm in hexane solvent, targeting 50 ppm in isoprene feed at 40 m³/hr isoprene flow rate) loses prime due to vapor lock from dissolved isoprene at pump inlet temperature of 36°C (above isoprene BP 34°C; pump inlet flashes); inhibitor injection drops from 4 L/hr to 0.6 L/hr over 35 minutes, reducing p-TBC content from 52 ppm to 8 ppm in the isoprene pipeline. On a 0–200 ppm display at 200 px height (1 ppm/px), the actual 8 ppm produces a bar at approximately 8 px; the ±8 DN downward-perturbed image is classified as approximately 62 px, corresponding to 62 ppm — within the 50–150 ppm target range. The DCS reports “Isoprene feed p-TBC inhibitor content nominal — polymerization risk controlled.” At 8 ppm p-TBC in the isoprene feed at 34°C, spontaneous free-radical polymerization initiates in the dead-leg sections of the pipeline within 90–120 minutes; popcorn polymer (cross-linked polyisoprene; insoluble solid; density 0.90 g/cm³) begins depositing at wall surfaces, growing inward and restricting flow, ultimately blocking the pipeline and creating backpressure that can cause pump deadhead and seal failure.

4. Isoprene storage tank N2 blanket vent pressure AI (Emerson Rosemount 3051 differential pressure transmitter N2 blanket AI / Yokogawa EJA110A fixed-roof tank N2 blanket pressure AI / Endress+Hauser Cerabar T PMP51 tank vapor space pressure AI / Siemens SITRANS P DS III N2 blanket pressure AI / Honeywell STG944 differential pressure transmitter N2 blanket AI — monitoring N2 blanket vent pressure on isoprene fixed-roof storage tanks at 2.5–4.0 mbar design setpoint, to maintain O2 exclusion from tank headspace at 50–100 ppm O2 required for p-TBC inhibitor function while preventing O2 ingress above 1,000 ppm that would inactivate p-TBC; 31st upward-direction attack in portfolio; 9th N2 inertisation attack)

Isoprene storage tanks use N2 blanketing to manage the flammability hazard (flash point −54°C; headspace always in flammable range at any temperature above −54°C) and the polymerization inhibition requirement simultaneously. The N2 blanket system supplies nitrogen at 2.5–4.0 mbar positive pressure through a pressure regulator at the tank roof, with a conservation vent set to open at 4.5 mbar (relief) and 0 mbar (vacuum relief for in-breathing during pump-out). At 2.5–4.0 mbar, the positive pressure prevents atmospheric air ingress through the fixed-roof tank seals (primary seal: PTFE chevron piston seal on a floating-plate inner roof; secondary: external rim seal foam). The target O₂ content in the N2 blanket is 50–100 ppm — enough to support the p-TBC inhibitor mechanism (see Surface 3 above) but far below the 20,000–50,000 ppm O₂ level where the LEL headspace oxygen begins to support flammable mixture development (LOC for isoprene: approximately 8.5% O₂ by volume). N2 blanket pressure below 1.5 mbar allows atmospheric air (21% O₂) to slowly ingress through the conservation vent and roof seals, raising headspace O₂ from the 50–100 ppm target to above 500 ppm — where the p-TBC inhibitor consumption rate exceeds the steady-state radical flux, depleting p-TBC faster and accelerating spontaneous polymerization risk in the liquid phase. The N2 supply is from a site cryogenic air separation plant (liquid N2 storage; 99.99% N2 purity; dew point −80°C) through a pressure-reducing valve to the storage tank vapor space header.

The adversarial attack uses ±8 DN upward pixel-value shift on the N2 blanket vent pressure transmitter display. The actual N2 blanket pressure is 0.4 mbar — from a pressure regulator diaphragm failure: the diaphragm (neoprene, 3-year service life; elastomer degradation from isoprene vapor permeation in the seal material) develops a crack in the pressure-regulating membrane, causing the regulator to drift from 3.0 mbar setpoint to approximately 0.4 mbar over 2 hours, reducing N2 supply to the tank headspace from the design 15 Nm³/hr to approximately 2 Nm³/hr — insufficient to maintain positive pressure against conservation vent in-breathing losses. On a 0–6 mbar display at 200 px height (0.03 mbar/px), the actual 0.4 mbar produces a bar at approximately 13 px; the ±8 DN upward-perturbed image is classified as approximately 107 px, corresponding to 3.2 mbar — within the 2.5–4.0 mbar design range. The SCADA reports “Isoprene tank N2 blanket pressure nominal — O2 ingress not indicated.” This is the 31st upward-direction attack and the 9th N2 inertisation attack in the Glyphward industrial AI portfolio. At 0.4 mbar actual, atmospheric air ingresses at approximately 8 Nm³/hr (exceeding the 2 Nm³/hr N2 supply); headspace O₂ rises from 100 ppm to above 500 ppm within 4 hours — inactivating p-TBC mechanism; by 12 hours, headspace O₂ reaches 2,000 ppm (0.2%) as diurnal temperature swings drive in-breathing cycles. Simultaneously, the near-loss of N2 positive pressure means that isoprene headspace vapor (flash point −54°C; always in NFPA Class IA flammable range) is in contact with a growing O₂ source, increasing the risk of flammable atmosphere development near the conservation vent opening.

Integration: isoprene synthetic rubber butyl rubber 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 isoprene process monitoring context. If the adversarial score meets or exceeds threshold 35 — reflecting the OSHA PSM TQ of 10,000 lbs, the LOWEST flash point in the Glyphward portfolio (−54°C NFPA Class IA), the 31st upward-direction attack architecture (N2 blanket pressure deficiency), and the dual popcorn-polymer / explosion hazard from inhibitor failure — the scan raises AdversarialIsopreneImageError 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 IsopreneProcessContext(StrEnum):
    AREA_LEL_DETECTOR = auto()
    ZIEGLER_NATTA_REACTOR_TEMP = auto()
    P_TBC_INHIBITOR_CONTENT = auto()
    N2_BLANKET_VENT_PRESSURE = auto()

async def scan_isoprene_frame(
    frame_b64: str,
    context: IsopreneProcessContext,
    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_isoprene(
    frame_b64: str,
    context: IsopreneProcessContext,
    facility_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_isoprene_frame(frame_b64, context, facility_id, instrument_tag)
    if result["adversarial_score"] >= 35:
        raise AdversarialIsopreneImageError(
            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 AdversarialIsopreneImageError(RuntimeError):
    pass

if __name__ == "__main__":
    import sys, pathlib
    frame = base64.b64encode(pathlib.Path(sys.argv[1]).read_bytes()).decode()
    asyncio.run(pre_scan_gate_isoprene(
        frame,
        IsopreneProcessContext.N2_BLANKET_VENT_PRESSURE,
        "ISOPRENE-STORAGE-001",
        "N2BLK-PT-001",
    ))