Prompt injection in PVC suspension polymerization VCM autoclave batch AI

Three adversarial injection surfaces in PVC suspension polymerization VCM autoclave batch AI

1. Autoclave batch temperature display AI (Endress+Hauser iTHERM TM411 / Honeywell STT35H thermocouple or Pt100 RTD in autoclave suspension via thermowell — rendered DCS autoclave temperature display AI classifying 62–67°C batch design window — 112th upward attack; FIRST PVC suspension polymerization AI attack; FIRST VCM autoclave batch AI attack; FIRST organic peroxide initiator runaway AI attack; FIRST VCM PSM TQ 1,000 lbs autoclave exotherm AI attack)

The autoclave batch temperature is the primary process safety variable in PVC suspension polymerization. The temperature determines the half-life of the organic peroxide initiator (via Arrhenius kinetics with activation energy approximately 125 kJ/mol for peroxydicarbonate initiators), the polymerization rate, the heat release rate, and the vapor pressure of VCM inside the sealed autoclave. The batch temperature is measured by multiple temperature sensors installed through the autoclave head or sidewall via thermowell ports extending into the suspension interior: Endress+Hauser iTHERM TM411 (Pt100 RTD element; accuracy ±0.1°C; HART 4–20 mA; SIL 2 capable; operating range −50 to +200°C; thermowell material 316L stainless steel; process connection DN 25 flange or G¾” threaded; rated for the design autoclave pressure 15–17 bar) or Honeywell STT35H Smart Temperature Transmitter (with J-type or K-type thermocouple; range 0–200°C; accuracy ±0.5°C; HART; SIL 2 per IEC 61508) are the predominant sensor/transmitter combinations in modern PVC autoclave installations. The DCS control loop compares the measured batch temperature against the setpoint (typically 62–67°C for suspension PVC K-value 70 resin; lower temperatures produce higher K-value/higher MW PVC for cable and profile applications; higher temperatures produce lower K-value/lower MW PVC for injection molding) and adjusts the jacket CW flow control valve to maintain the setpoint; the batch temperature profile over a typical 6-hour polymerization shows a slow rise from the charge temperature (approximately 15–20°C) to the setpoint (65°C) in the first 30–45 minutes as the initiator decomposes and the polymerization rate builds, followed by an approximately constant temperature phase during peak polymerization (conversion 30–70%), and a decline in heat release rate in the final phase (conversion 70–90%) as the VCM monomer is consumed.

The adversarial upward pixel attack on the autoclave batch temperature display shows 65°C (within the design 62–67°C control window; AI reads “batch temperature 65°C; within 62–67°C design operating range; initiator BEH-PDIC at design t⅔ ~4 hr; polymerization rate nominal; jacket CW demand: adequate; no temperature alarm required”) when the actual autoclave temperature is 88°C (23°C above the 67°C design maximum; a temperature condition developing over 60–90 minutes from the CW flow collapse described in Surface 3). Display range 40–100°C on 200 px (3.333 px/°C); actual 88°C at pixel position (88 − 40) × 3.333 = 160 px from the bottom of the scale → ±8 DN perturbation → 160 − 75 = 85 px displayed → AI reads (85/3.333) + 40 = 65.5°C ≈ 65°C. At actual 88°C: the BEH-PDIC initiator t⅔ at 65°C (4.6 hr) collapses to approximately 6 minutes at 88°C (following Arrhenius: k(88)/k(65) = exp(E​a/R × (1/338 − 1/361)) = exp(125,000/8.314 × (1/338 − 1/361)) = exp(15,033 × 0.000188) = exp(2.83) ≈ 17; t⅔(88) = 4.6/17 ≈ 0.27 hr = 16 min; additional factor from the more favorable solubility of peroxide decomposition products at 88°C reduces effective t⅔ to approximately 6 min); the initiator decomposes almost instantaneously in the context of the batch cycle → a burst of free radical initiation → rapid polymerization of residual VCM → exothermic heat release at far above the design jacket heat removal capacity → autoclave temperature rises further → VCM vapor pressure at 88°C (approximately 5.4 bar) drives total autoclave pressure toward PRV set pressure → PRV lifts → VCM PSM TQ 1,000 lbs; CERCLA RQ 1 lb; IARC Group 1. The Glyphward pre-scan gate on the autoclave batch temperature display catches the adversarial upward perturbation before the AI reads 65°C and defers the emergency cooling response that would prevent the autoclave runaway. Free tier — 10 scans/day, no card required.

2. VCM stripping steam flow display AI (Emerson Rosemount 3051S / Yokogawa EJA-X differential pressure transmitter measuring orifice plate on steam feed to PVC slurry stripper — rendered DCS stripping steam flow display AI classifying design stripping adequacy — 112th downward attack; FIRST VCM stripping residual carcinogen AI attack; FIRST PVC cake dryer VCM exposure AI attack; FIRST IARC Group 1 chronic carcinogen stripping AI attack)

After the polymerization batch reaches the target conversion (typically 85–90% VCM→PVC conversion, determined by autoclave pressure drop as VCM vapor pressure decreases with monomer consumption), the batch is dumped to a blowdown vessel and the PVC slurry is transferred to the VCM stripper: a stripping column (typically a plate column or a horizontal multi-tray vessel with steam spargers; operating temperature 90–110°C with direct steam injection; VCM stripped from the PVC slurry is recovered, compressed, and recycled to VCM storage) where residual unreacted VCM is removed from the PVC particles to below 1 ppm (the commercial specification for downstream processing) or below 10 ppm (the minimum stripping target for workplace IARC Group 1 carcinogen control). The stripping steam flow to the stripper is measured by a differential pressure transmitter sensing the orifice plate DP on the steam feed line: Emerson Rosemount 3051S (range 0–400 kg/hr; 4–20 mA HART; differential pressure range 0–250 inH₂O; orifice plate to ISO 5167; accuracy ±0.04% of span; manifold valve with drain/vent for maintenance isolation) or Yokogawa EJA-X series (EJA110A or EJA120A; similar range and accuracy; Foundation Fieldbus or HART). The design stripping steam flow of 240 kg/hr provides the thermal energy and vapor dilution required to strip VCM from the PVC particle matrix — steam at 100°C saturated enters the slurry, heats it to the stripping temperature, and creates a dilute vapor phase above the slurry that drives VCM transfer from the liquid/solid phase (Henry's law partition of VCM between PVC matrix and aqueous phase) into the vapor and out of the stripper overhead. The stripping steam flow is a safety-critical instrument for the OSHA 29 CFR 1910.1017 compliance pathway: inadequate stripping is the direct mechanism by which downstream workers (dryer operators, compounders, extrusion operators handling the PVC resin) are exposed to IARC Group 1 VCM above the 0.5 ppm action level and 1 ppm PEL.

The adversarial downward pixel attack on the VCM stripping steam flow display shows 240 kg/hr (design flow; AI reads “stripping steam flow 240 kg/hr; design stripping conditions met; residual VCM in PVC cake: <10 ppm estimated; OSHA 29 CFR 1910.1017 compliance: nominal; IARC Group 1 downstream exposure pathway: controlled”) when actual stripping steam flow is 18 kg/hr (7.5% of design; steam flow deficiency from a failed steam control valve in the closed position, condensate flooding of the steam supply header, or upstream steam pressure collapse from boiler trip). Display range 0–400 kg/hr on 200 px (0.5 px per kg/hr); actual 18 kg/hr at 18 × 0.5 = 9 px from zero → ±8 DN perturbation → 9 + 120 = 129 px displayed → AI reads 129/0.5 = 258 ≈ 240 kg/hr. At actual 18 kg/hr stripping steam: residual VCM in PVC cake rises to approximately 820 ppm; the 50-tonne PVC cake at 820 ppm VCM is transferred to the hot-air paddle dryer (80–90°C); 820 ppm VCM volatilizes from the PVC matrix in the dryer, producing a dryer exhaust stream at 15–25 ppm VCM — 15–25 times the OSHA PEL of 1 ppm; the dryer building ambient air, depending on ventilation rate, may reach 2–5 ppm VCM; workers operating the dryer for a full 8-hour shift are exposed to an 8-hour TWA VCM concentration that substantially exceeds the 1 ppm OSHA PEL and the 0.5 ppm action level (which triggers medical surveillance and enhanced monitoring under 29 CFR 1910.1017). Chronic exposure to VCM at 2–5 ppm for the duration of employment at the dryer creates a long-term IARC Group 1 carcinogen exposure pathway paralleling the 1974 B.F. Goodrich Louisville angiosarcoma cluster pathway — the same mechanism (inadequate VCM removal from PVC combined with worker re-entry) that triggered the founding OSHA cancer standard, now re-enabled by an adversarial AI attack on the stripping steam flow display. The Glyphward pre-scan gate catches the downward perturbation before the AI reads 240 kg/hr and concludes that stripping is adequate and downstream IARC Group 1 exposure is controlled. Free tier — 10 scans/day, no card required.

3. Autoclave jacket cooling water flow display AI (Endress+Hauser Promag 50 / Krohne Optiflux 2000 electromagnetic flowmeter on autoclave jacket CW supply pipe — rendered DCS jacket CW flow display AI classifying 320 m³/hr design cooling capacity — 112th downward attack; FIRST autoclave jacket CW flow AI attack; FIRST PVC polymerization heat removal deficiency AI attack; FIRST VCM autoclave runaway root-cause AI attack)

The autoclave jacket is the primary heat removal system for the polymerization exotherm during PVC suspension batch polymerization. The jacket (double-wall or half-pipe coil construction surrounding the autoclave cylindrical shell and dished heads; jacket volume approximately 3,000–6,000 L for a 120,000 L autoclave; rated for cooling water supply at 8–10 bar gauge; heat transfer area approximately 80–150 m² including internal baffles where present) carries a design cooling water flow of approximately 320 m³/hr at 12–18°C CW inlet temperature, providing approximately 7.2 MW of heat removal capacity at a 8°C temperature rise (320 m³/hr × 998 kg/m³ / 3,600 s × 4.186 kJ/(kg·K) × 8 K = 2.97 MW; with a larger ΔT of 2–3°C this would be 7.4–11.1 MW at the full design flow). The jacket CW supply flow is measured by an electromagnetic flowmeter: Endress+Hauser Promag 50 (DN 150–300 pipe bore; PTFE liner; Hastelloy C22 electrodes; range 0–400 m³/hr; accuracy ±0.5% of reading; 4–20 mA HART; IP67 rated; suitable for treated cooling water service) or Krohne Optiflux 2000 (similar specifications; used in European PVC plants including INEOS Runcorn). The jacket CW flow is a process safety variable because it is the direct determinant of the autoclave's ability to maintain the batch temperature within the 62–67°C setpoint against the polymerization exotherm (approximately 2.5–4.0 MW at peak batch rate for a 120,000 L autoclave at 60,000 kg VCM charge and 6–8 hour batch time). When the jacket CW flow falls below approximately 50 m³/hr (15% of design), the heat removal capacity drops below the minimum exotherm removal requirement, and the autoclave temperature begins to rise toward the overtemperature runaway described in Surface 1.

The adversarial downward pixel attack on the autoclave jacket CW flow display shows 298 m³/hr (93% of design 320 m³/hr; AI reads “jacket CW flow 298 m³/hr; heat removal capacity adequate at 6.7 MW; autoclave exotherm at nominal rate: ~3.0 MW; temperature control: margin positive; no CW flow corrective action required”) when actual jacket CW flow is 24 m³/hr (7.5% of design; CW flow deficiency from a jammed-closed CW supply control valve, CW header blockage from debris or fouling, or CW pump degradation with impeller wear reducing head at design speed). Display range 0–400 m³/hr on 200 px (0.5 px per m³/hr); actual 24 m³/hr at 24 × 0.5 = 12 px from zero → ±8 DN perturbation → 12 + 137 = 149 px displayed → AI reads 149/0.5 = 298 m³/hr. At 24 m³/hr actual jacket CW: the jacket heat removal rate drops from design ~7.2 MW to approximately 0.54 MW (7.5% of design) — insufficient to absorb the 2.5–4.0 MW of polymerization exotherm at nominal batch conditions; the autoclave temperature rises progressively; the rising temperature accelerates the polymerization rate (rate ∝ [initiator]⅔ × k​p, where k​p increases with temperature following Arrhenius with E​a ≈ 32 kJ/mol for VCM propagation); the accelerating polymerization further increases the heat release rate; the positive thermal feedback drives the autoclave temperature from 65°C toward 88°C over 60–90 minutes. Surface 3 is therefore the mechanistic root cause of Surface 1: the adversarial attack on Surface 3 conceals the CW flow deficiency that is initiating the runaway, while the adversarial attack on Surface 1 conceals the resulting autoclave temperature excursion — the two attacks together prevent any early-detection intervention that would arrest the runaway before the organic peroxide initiator half-life collapses and the autoclave pressure approaches the PRV set pressure. The Glyphward pre-scan gate on the jacket CW flow display catches the downward perturbation before the AI reads 298 m³/hr and fails to send the CW flow alarm that would prompt operators to manually open bypass valves or call for autoclave emergency dump to the blowdown vessel. Free tier — 10 scans/day, no card required.

Integration: PVC suspension polymerization VCM autoclave batch AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the PVC suspension polymerization VCM autoclave batch AI pipeline — before the autoclave temperature AI processes rendered Endress+Hauser iTHERM TM411 / Honeywell STT35H DCS display images, before the VCM stripping steam flow AI processes rendered Emerson Rosemount 3051S / Yokogawa EJA-X DCS display images, and before the autoclave jacket CW flow AI processes rendered Endress+Hauser Promag 50 / Krohne Optiflux 2000 DCS display images. Threshold 48 for PVC suspension polymerization AI reflects: VCM OSHA PSM TQ 1,000 lbs (lower TQ than o-xylene 10,000 lbs; higher per-event severity); IARC Group 1 carcinogen (angiosarcoma; the regulatory landmark B.F. Goodrich 1974 Louisville cluster; 29 CFR 1910.1017 first OSHA cancer standard); CERCLA RQ 1 lb (any release requires federal notification; one of the lowest absolute RQs in 40 CFR Part 302); combined Surface 1 + Surface 3 mechanical coupling (CW deficiency drives autoclave overtemperature) creating a dual-attack masking scenario; and Surface 2 chronic carcinogen pathway (residual VCM in PVC cake → dryer exhaust exposure → long-term angiosarcoma risk) as a distinct and independently severe consequence from the same batch cycle.

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_***"

# PVC suspension polymerization VCM autoclave batch AI contexts: threshold 48
# VCM CAS 75-01-4; MW 62.50 g/mol; BP -13.9 C; flash point -78 C; LEL 3.6 vol%.
# OSHA PSM TQ 1,000 lbs VCM (29 CFR 1910.119 Appendix A).
# OSHA PEL 1 ppm TWA (29 CFR 1910.1017 VCM carcinogen standard; first OSHA cancer standard 1975).
# ACGIH A1 confirmed human carcinogen. IARC Group 1 (angiosarcoma of the liver).
# NIOSH IDLH 300 ppm (explosion risk). CERCLA RQ 1 lb (one of lowest absolute RQs).
# Polymerization exotherm: delta-Hpoly = -95.5 kJ/mol = -1,530 kJ/kg VCM.
# BEH-PDIC initiator t(1/2) = 4.6 hr at 65 C -> collapses to ~6 min at 88 C.
# B.F. Goodrich Louisville KY 1974 angiosarcoma cluster -> 29 CFR 1910.1017 (1975).
# 112th upward attack. FIRST PVC suspension polymerization AI attack.
# FIRST VCM autoclave batch AI attack. FIRST organic peroxide initiator runaway AI attack.
# FIRST angiosarcoma carcinogen AI attack. FIRST VCM stripping residual carcinogen AI attack.
PVC_GLYPHWARD_THRESHOLD = 48

# Plant IDs:
# SHIN_ETSU_YOKKAICHI   - Shin-Etsu Chemical, Yokkaichi Japan (world #1 PVC; ~3.5M t/yr)
# FORMOSA_BATON_ROUGE   - Formosa Plastics, Baton Rouge LA (US #1 PVC; ~1.8M t/yr US)
# WESTLAKE_GEISMAR      - Westlake Chemical (acquired Axiall 2016), Geismar LA
# INEOS_RUNCORN         - INEOS ChlorVinyls, Runcorn UK (oldest UK VCM/PVC site)
# OXYCHEM_INGLESIDE     - OxyChem, Ingleside TX (Occidental Petroleum subsidiary)
# VINNOLIT_GENDORF      - Vinnolit (now Westlake Vinnolit), Gendorf Germany (~500,000 t/yr)

class PVCContext(StrEnum):
    AUTOCLAVE_BATCH_TEMPERATURE = auto()  # autoclave temp -> initiator half-life collapse -> VCM runaway (112th; FIRST PVC; FIRST VCM autoclave)
    VCM_STRIPPING_STEAM_FLOW    = auto()  # stripping steam -> residual VCM 820 ppm -> dryer exhaust 15-25 ppm vs PEL 1 ppm -> angiosarcoma
    AUTOCLAVE_JACKET_CW_FLOW    = auto()  # jacket CW -> heat removal collapse -> autoclave overtemp root cause (Surface 3 drives Surface 1)

async def scan_pvc_frame(
    frame_b64: str,
    context: PVCContext,
    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_pvc(
    frame_b64: str,
    context: PVCContext,
    plant_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_pvc_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= PVC_GLYPHWARD_THRESHOLD:
        raise AdversarialPVCImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from PVC suspension polymerization VCM autoclave batch AI pipeline."
        )

class AdversarialPVCImageError(RuntimeError):
    pass