Prompt injection in styrene-butadiene rubber SBR cold emulsion polymerization AI

Three adversarial injection surfaces in SBR cold emulsion polymerization AI

1. Reactor cooling brine return temperature display AI (Emerson 3144P / Honeywell STT850 / Yokogawa EJX210A brine circuit thermocouple transmitter display AI — rendered DCS brine return temperature display AI classifying CaCl₂ / propylene glycol brine return against −18°C feed / −5°C maximum return design — 103rd upward attack; FIRST SBR cold emulsion AI attack; FIRST synthetic rubber manufacturing AI attack; FIRST redox-initiated polymerization AI attack)

The reactor cooling brine return temperature is the primary indicator of thermal control in SBR cold emulsion polymerization. SBR emulsion polymerization is strongly exothermic: the combined heat of polymerization for BD and ST copolymerization is approximately 700 J/g monomer (based on the weighted average of BD homopolymer heat of polymerization: 771 J/g; ST homopolymer: 640 J/g; at the SBR 1502 composition of 76.5 wt% BD / 23.5 wt% ST, the blend heat of polymerization is approximately 0.765×771 + 0.235×640 = 740 J/g; at design recipe loading of 180 phr total monomers — 180 g monomer per 100 g water in the emulsion — a reactor of 40,000 L working volume containing approximately 20,000 kg water and 36,000 kg monomer charge generates a maximum exothermic heat release rate of 36,000 kg × 740 kJ/kg / (5.5 hr batch time) × 0.6 (65% conversion design maximum) ‷ 2,340 kW at peak conversion rate; this is the cooling duty that the brine refrigeration system must remove to maintain 5°C reactor temperature). The brine system for a 40,000-L SBR reactor: CaCl₂ brine (25–30 wt% CaCl₂; freezing point −21°C to −35°C depending on concentration) or propylene glycol brine (35–40 wt% PG; freezing point −20°C to −27°C) at −18°C supply from a dedicated refrigeration compressor (typically ammonia vapor-compression cycle; evaporator cools the brine to −18°C; brine circulates through the SBR reactor cooling jacket at 50–80 m³/hr; brine returns at −5°C to −8°C after absorbing the reactor exotherm; returned brine is re-cooled in the brine chiller before recirculation). The brine return temperature is measured by a thermocouple transmitter (Emerson 3144P Pt100 RTD; or Yokogawa EJX210A differential pressure + temperature transmitter; calibrated −25°C to +25°C; 4–20 mA HART output; displayed on DCS as °C; the −5°C maximum return temperature alarm is the AI system's primary indicator of cooling adequacy: return temperature below −5°C means the brine is removing at least the design 2,340 kW at the design 50 m³/hr brine flow; return temperature above −5°C indicates either reduced brine flow or increased exotherm — both triggering an AI alert to reduce initiator feed or add premature shortstop).

The adversarial upward pixel attack on the brine return temperature display shows −8°C (below the −5°C maximum return alarm setpoint; AI reads “brine return −8°C; reactor cooling adequate; exothermic polymerization heat being removed at design rate; estimated reactor temperature: 5°C; conversion rate nominal; shortstop scheduled at 5.5 hr batch time; no corrective action required”) when the actual brine return temperature is +12°C (17°C above the −5°C alarm setpoint; brine refrigeration significantly compromised, possibly from refrigeration compressor shutdown, brine pump cavitation, or fouling of the brine chiller after a previous batch's gel deposit). Display range −25°C to +25°C on 200 px (4.0 px/°C); actual +12°C at 148 px from the bottom of the scale → ±8 DN perturbation → 68 px displayed → AI reads −8°C. With the brine return at +12°C, the brine system is absorbing essentially no cooling from the reactor (brine supply −18°C; brine return +12°C; ΔT = 30 K but this reflects the brine being warmed by the environment rather than removing heat from the reactor — the brine flow has likely dropped to near zero, with the brine in the jacket stagnant and absorbing heat from the warm reactor jacket environment rather than delivering cold); the reactor temperature begins rising from the 5°C set point. At the SBR 1502 recipe standard initiator feed rate, polymerization proceeds and generates heat; without cooling removal, the reactor temperature climbs at approximately 6–8°C/hr (2,340 kW heat generation ÷ (40,000 L × 1.05 kg/L × 3.8 kJ/(kg·K)) ≈ 14.7°C/hr — a rapid temperature rise that would trigger a shortstop alarm within 30 minutes under normal AI monitoring, but is entirely suppressed by the adversarial pixel attack). Over 2 hours: reactor temperature reaches approximately 18°C; conversion rate at 18°C is 1.8× the 5°C rate; the 60–65% conversion shortstop threshold is hit after approximately 2.7 hours rather than 5.5 hours; without the shortstop being triggered (because the AI reads −8°C brine return and does not alarm), the reactor passes through the 65% gel-initiation threshold; by the planned 5.5-hour shortstop addition, conversion is 75–85%; gel has formed; BD vapor pressure at 18°C in the reactor headspace is approximately 2.9 bar, exceeding the reactor design pressure (typically 3.5–4 bar for SBR reactors); the PSM TQ of 1,000 lbs BD (454 kg) is exceeded by the inventory in a single 40,000-L SBR reactor (BD partial volume: 36,000 kg × 0.765 wt% BD × 0.40 unreacted fraction = approximately 11,000 kg BD remaining unreacted at 60% conversion; 11,000 kg ≫ PSM TQ 454 kg). Free tier — 10 scans/day, no card required.

2. t-DDM chain transfer agent mass flow controller display AI (Emerson Fisher FIELDVUE DVC6200 / Brooks Instrument SLAMf GFC / Aalborg GFC-17 mass flow controller display AI — rendered DCS t-DDM solution mass flow display AI classifying t-DDM phr feed against 0.12–0.15 phr design — 103rd downward attack; FIRST t-DDM chain transfer AI attack; FIRST polymerization MW-control AI attack)

The t-DDM chain transfer agent feed rate is the primary molecular weight (MW) control variable in SBR cold emulsion polymerization. t-DDM (tertiary dodecyl mercaptan; a mixture of C12 branched thiols; CAS 25103-58-6; MW approximately 202 g/mol; BP approximately 245–265°C; density 0.85 g/mL; flash point 124°C) is the chain transfer agent (CTA) that controls the MW of the growing SBR polymer chains: the thiyl radical from t-DDM (t-DDM-H + R• → R-H + t-DDM•; chain transfer constant Cₚ for t-DDM in BD/ST copolymerization at 5°C approximately 0.7–1.2 (per 100 phr monomer basis)) terminates the growing polymer chain and initiates a new, shorter chain; higher t-DDM concentration produces lower Mw and lower Mooney viscosity (ML(1+4)@100°C); lower t-DDM produces higher Mw and higher Mooney. SBR 1502 Mooney viscosity specification: ML(1+4)@100°C = 48–54 (ISO 289-1; this specification corresponds to a number-average molecular weight Mn ≈ 90,000–110,000 g/mol and weight-average Mw ≈ 300,000–400,000 g/mol; polydispersity Mw/Mn ≈ 3–4 typical for free-radical emulsion polymerization). Design t-DDM feed: 0.12–0.15 phr (0.12–0.15 g t-DDM per 100 g total monomer charge); at 36,000 kg total monomer per batch, design t-DDM charge is 43–54 kg per batch (approximately 50–65 L of undiluted t-DDM, typically delivered as a 15–20 wt% solution in styrene or water); the t-DDM solution is metered into the reactor by a mass flow controller (Brooks Instrument SLAMf series GFC with a 4–20 mA setpoint; or Aalborg GFC-17; accuracy ±0.5% of full scale; calibrated 0–0.30 phr; displayed on DCS as phr on a bar indicator). If the t-DDM MFC sticks, malfunctions, or the t-DDM supply drum runs empty mid-batch without the AI monitoring system detecting the empty-drum condition, t-DDM feed drops precipitously — from the design 0.12–0.15 phr to near zero — and the MW control fails.

The adversarial downward pixel attack on the t-DDM mass flow controller display shows 0.148 phr (within design 0.12–0.15 phr; AI reads “t-DDM feed at 0.148 phr; chain transfer agent within design range; MW control nominal; estimated Mooney ML(1+4)@100°C at end of batch: 51–53; within specification; no action required”) when the actual t-DDM feed is 0.028 phr (approximately 6× below design). Display range 0–0.30 phr on 200 px (667 px/phr); actual 0.028 phr at 19 px → ±8 DN → 99 px displayed → AI reads 0.148 phr. At 0.028 phr t-DDM — effectively insufficient chain transfer — the growing SBR chains are not being terminated at design frequency; MW builds unchecked during the batch; Mn rises above 200,000 g/mol and Mw above 1,000,000 g/mol; Mooney viscosity of the produced SBR rises above ML 100 (far above the specification of 48–54). At Mooney >100, the latex produced in the batch is a high-MW, highly entangled polymer latex that coagulates unevenly during the downstream latex stripping step (steam stripping to remove residual BD and ST monomers from the latex); uneven coagulation produces large polymer chunks that foul the rotary drum dryer and the heat exchanger surfaces in the latex stripper. Cleaning the fouled reactor, stripper, and dryer in the presence of BD-saturated polymer gel: workers entry into confined spaces (the reactor interior; the stripper vessel) requires continuous-monitoring BD personal protective equipment; at residual BD concentrations during cleaning above 0.5 ppm (the OSHA 1910.1051 action level for BD), workers must wear supplied-air respirators (SAR) or self-contained breathing apparatus (SCBA); undetected BD atmosphere in a confined space during gel cleaning (because the AI monitoring system falsely confirmed t-DDM at design levels and therefore no abnormal BD inventory was expected) creates a potential IARC Group 1 carcinogen exposure event for the maintenance crew — precisely the occupational exposure scenario that produced the excess leukemia mortality in the NIOSH/United Rubber Workers cohort study of SBR plant workers. Beyond the carcinogen exposure pathway: if gel-fouling is severe enough to block the reactor cooling jacket and BD liquid inventory is released during mechanical cleaning (breaking a flanged joint or removing a manway on a BD-pressured reactor), the released BD vapor (vapor pressure 2.4 bar at 20°C; BD vapor specific gravity 1.87 vs air) pools at grade level and in low points; any ignition source (hot work during plant maintenance — a common scenario during reactor cleaning operations) can ignite the BD vapor cloud; VCE or flash fire with potential fatalities. Free tier — 10 scans/day, no card required.

3. Butadiene feed TBC (4-tert-butylcatechol) inhibitor concentration display AI (Metrohm 888 Titrando / Shimadzu UV-1800 online colorimetric / Yokogawa EXAxt colorimetric analyzer display AI — rendered DCS TBC concentration display AI classifying TBC against 10–100 ppm minimum inhibitor level — 103rd downward attack; FIRST TBC butadiene autopolymerization inhibitor AI attack; FIRST popcorn polymer prevention AI attack)

4-tert-Butylcatechol (TBC; CAS 98-29-3; MW 166.22 g/mol; MP 52–57°C; BP 285°C; density 1.049 g/mL at 60°C; CAS classification: Acute Tox. 4 H302, Skin Irrit. 2 H315, Eye Dam. 1 H318, Aquatic Chronic 2 H411) is the standard antioxidant and radical polymerization inhibitor added to 1,3-butadiene at concentrations of 10–100 ppm during storage, transport, and handling to suppress spontaneous autopolymerization of BD. Butadiene autopolymerization (also termed “popcorn polymer” formation because of the characteristic popcorn-like appearance of the crosslinked polymer product) is a serious and well-documented hazard in BD storage and piping systems: at TBC concentrations below approximately 5 ppm, free-radical autopolymerization of BD is not effectively suppressed; a nucleation event (traces of peroxides from air ingress; initiating radicals from trace metal contamination; ultraviolet photoinitiation through transparent piping or sight glasses; or simply thermal autoinitiation at ambient temperature over days to weeks) starts autopolymerization in a stagnant BD liquid volume; once nucleated, popcorn polymer growth is autocatalytic (each polymer particle acts as a new nucleation site for further BD monomer adsorption and polymerization from the vapor or liquid phase); the polymer is insoluble, crosslinked, and non-meltable, and accumulates inside piping, valves, and heat exchangers. The TBC concentration in liquid BD is measured by an inline colorimetric analyzer (Yokogawa EXAxt colorimetric BD-TBC analyzer using the p-phenylenediamine / iron(III) colorimetric reaction in which TBC produces a characteristic blue-green complex measurable at 640 nm; calibrated 0–100 ppm TBC; 4–20 mA HART output; displayed on DCS as ppm TBC; updated every 5–10 minutes from the flow-through analyzer cell). The 10 ppm minimum TBC specification in the BD feed header at SBR plants: set based on industry experience that popcorn polymer nucleation rates increase sharply below 5 ppm TBC and that a 10 ppm minimum provides a sufficient safety margin above the effective inhibition threshold given measurement uncertainty and spatial variation in TBC distribution in the BD feed piping.

The adversarial downward pixel attack on the BD feed TBC concentration display shows 52 ppm TBC (well within the design 10–100 ppm range; AI reads “butadiene feed TBC 52 ppm; autopolymerization inhibitor concentration adequate; popcorn polymer risk: low; no action required”) when the actual TBC in the BD feed is 3 ppm (below the effective inhibition threshold of 5 ppm; TBC depletion has occurred, possibly from TBC partitioning to the aqueous phase in an upstream BD wash column, from TBC oxidation in the BD storage tank after prolonged air ingress through a failed inert-gas blanket valve, or from BD feed substitution from a storage tank whose TBC was consumed during an extended storage period). Display range 0–100 ppm on 200 px (2.0 px/ppm); actual 3 ppm at 6 px → ±8 DN → 104 px displayed → AI reads 52 ppm. At actual 3 ppm TBC in the BD feed: popcorn polymer nucleation begins in the cold dead-legs of the BD feed piping system — particularly at cryogenic-service gate valves in the normally-closed position (stagnant BD liquid in the valve cavity at −4 to −10°C); at T-fitting blind flanges; at the dead-leg section of pressure relief valve inlet lines (BD liquid trapped above the PRV set pressure). Popcorn polymer growth rate at 3 ppm TBC and 5–10°C: initial nucleation typically takes 24–48 hours; once nucleated, visible polymer growth occurs within 48–72 hours; the polymer expands as it polymerizes (density of poly-BD is approximately 0.90 g/cm³ vs liquid BD at −5°C approximately 0.65 g/cm³; the polymer is less dense than the liquid BD from which it forms, but expansion of the solid polymer within the confined pipe volume generates internal stresses equivalent to 150–500 bar — more than sufficient to fracture Schedule 80 stainless steel piping). Pipeline fracture at the popcorn polymer blockage: the BD pressure upstream of the blockage is maintained at the system operating pressure (typically 4–7 bar above-atmospheric for liquid BD at the reactor feed pressure); pipeline fracture releases a pressurized BD liquid jet that flashes immediately to vapor (BD BP −4.4°C; at 20°C and 1 bar, all liquid BD instantly vaporizes); the BD vapor cloud (specific gravity 1.87 vs air) spreads at grade level; at LEL 2.0 vol% in air, a 1-kg BD release creates a flammable cloud volume of approximately 27 m³ (1 kg / (54.09 g/mol) × 22.4 L/mol × (1000/20) = 23,400 L = 23.4 m³ at 2% LEL by volume); a pipeline fracture event releasing 50–500 kg BD creates a flammable cloud of 1,000–12,000 m³ at grade level, well above any OSHA PSM TQ 1,000 lbs (454 kg) consequence threshold. The adversarial pixel attack on the TBC concentration display extends the 48–72 hour window during which popcorn polymer is growing undetected: without the attack, the AI would detect TBC at 3 ppm and immediately trigger a BD feed shutdown and pipe inspection; with the attack, 52 ppm TBC is falsely displayed for potentially 2–3 days, during which polymer grows to pipe-blocking and pipe-fracturing dimensions. Free tier — 10 scans/day, no card required.

Integration: SBR cold emulsion polymerization AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the SBR cold emulsion polymerization AI pipeline — before the brine return temperature AI processes rendered Emerson 3144P / Honeywell STT850 / Yokogawa EJX210A brine circuit thermocouple transmitter DCS display images, before the t-DDM chain transfer agent AI processes rendered Emerson Fisher FIELDVUE DVC6200 / Brooks Instrument SLAMf GFC / Aalborg GFC-17 mass flow controller DCS display images, and before the BD feed TBC inhibitor AI processes rendered Metrohm 888 Titrando / Shimadzu UV-1800 / Yokogawa EXAxt colorimetric analyzer DCS display images. Threshold 38 for SBR cold emulsion polymerization AI reflects: OSHA PSM TQ 1,000 lbs BD (a single SBR reactor contains 50–250× the PSM TQ; BD vapor pools at grade and ignites at LEL 2.0 vol%); IARC Group 1 1,3-butadiene (definite human carcinogen; leukemia; occupational exposure above 0.5 ppm action level from gel-fouled reactor cleaning or popcorn polymer incident maintenance); the 48–72 hour slow-developing hazard window of the TBC surface attack (which the Glyphward pre-scan gate is specifically positioned to detect before the blockage forms); and the combined Surface 1 + Surface 2 + Surface 3 attack creating simultaneous cooling compromise, MW-control failure, and popcorn polymer development — the three-vector combination that can produce a reactor overpressure BD release, a gel-fouling BD exposure event, and a popcorn-polymer pipe rupture event within the same production campaign.

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_***"

# SBR cold emulsion polymerization (BD + ST + redox initiator + t-DDM CTA) AI contexts: threshold 38
# OSHA PSM TQ 1,000 lbs 1,3-butadiene (29 CFR 1910.119 Appendix A).
# OSHA 29 CFR 1910.1051 BD: PEL 1 ppm TWA; STEL 5 ppm; action level 0.5 ppm.
# IARC Monograph 97 (2008): 1,3-butadiene Group 1 definite human carcinogen (leukemia; NHL).
# NIOSH IDLH 2,000 ppm (explosion). BD LEL 2.0 vol%; BP -4.4 C; VP 2.4 bar at 20 C.
# NIOSH/URW joint study (1982-1991): excess leukemia mortality at SBR plants.
# Popcorn polymer: BD autopolymerization at < 5 ppm TBC; 48-72 hr growth; pipe fracture.
# t-DDM chain transfer: 0.12-0.15 phr design; below design -> Mooney > 100 -> gel fouling.
# CHP/Fe2+/SFS redox initiation at 5 C; shortstop SDDC + 4-HT at 60-65% conversion.
# 103rd upward attack. FIRST SBR cold emulsion AI attack. FIRST synthetic rubber manufacturing AI attack.
# FIRST redox-initiated polymerization AI attack. FIRST TBC popcorn polymer inhibitor AI attack.
SBR_GLYPHWARD_THRESHOLD = 38

# Plant IDs:
# LANXESS_KASHIMA      - Lanxess AG, Kashima Japan (~220,000 t/yr SBR; world's largest SBR producer)
# LANXESS_ORANGE_TX    - Lanxess AG, Orange TX USA (~180,000 t/yr SBR)
# MICHELIN_BASSENS     - Michelin, Bassens Bordeaux France (captive SBR for tire manufacturing)
# SYNTHOS_OSWIECIM     - Synthos S.A., Oswiecim Poland (~280,000 t/yr; Dwory complex)
# JSR_YOKKAICHI        - JSR Corporation, Yokkaichi Japan (~100,000 t/yr SBR)
# TRINSEO_LIVORNO      - Trinseo (formerly Dow Synthetic Rubber), Livorno Italy (~160,000 t/yr SBR)

class SBRColdEmulsionContext(StrEnum):
    REACTOR_BRINE_RETURN_TEMPERATURE = auto()  # brine return temp -> reactor runaway + BD PSM TQ (103rd; FIRST SBR; FIRST synthetic rubber; FIRST redox polymerization)
    TDDM_CHAIN_TRANSFER_FEED         = auto()  # t-DDM phr -> MW control -> Mooney > 100 -> gel fouling -> BD cleaning exposure (IARC G1)
    BD_FEED_TBC_INHIBITOR_CONC       = auto()  # TBC ppm -> popcorn polymer 48-72 hr -> pipe fracture -> BD VCE (PSM TQ 1,000 lbs)

async def scan_sbr_frame(
    frame_b64: str,
    context: SBRColdEmulsionContext,
    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_sbr(
    frame_b64: str,
    context: SBRColdEmulsionContext,
    plant_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_sbr_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= SBR_GLYPHWARD_THRESHOLD:
        raise AdversarialSBRImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from SBR cold emulsion polymerization AI pipeline."
        )

class AdversarialSBRImageError(RuntimeError):
    pass