Prompt injection in MEKP methyl ethyl ketone peroxide organic peroxide manufacturing AI

Three adversarial injection surfaces in MEKP organic peroxide manufacturing AI

1. Reaction vessel jacket cooling water supply temperature display AI (Honeywell TDC 3000 MEKP reactor jacket cooling water supply temperature AI / Yokogawa CENTUM VP organic peroxide batch reactor jacket cooling AI / Emerson DeltaV MEKP peroxidation vessel jacket temperature AI / ABB 800xA MEKP reaction vessel heat removal AI / Rosemount 3244MV MEKP batch reactor jacket cooling supply temperature AI — rendered DCS temperature display AI classifying the chilled water (CW) supply temperature to the MEKP peroxidation reactor jacket against the 2–8°C design range ensuring heat removal from the exothermic MEK + H2O2 peroxidation reaction exceeds heat generation at all conversion rates, maintaining reactor contents below 20°C and well below the SADT of 40°C; 59th upward-direction attack — FIRST MEKP/methyl ethyl ketone peroxide manufacturing attack; FIRST commercial organic peroxide manufacturing attack; FIRST SADT-zone reaction vessel temperature upward attack)

MEKP is manufactured in dedicated batch or semi-continuous reactors (glass-lined jacketed vessels; ASME Section VIII; typical capacity 1–10 m³; operating at atmospheric pressure; agitator speed 60–120 rpm; temperature measured by PT-100 RTD in thermowell). The peroxidation reaction — MEK + H2O2 + H2SO4 catalyst at 5–15°C — is mildly but continuously exothermic: ΔH for MEK peroxidation ≈ −18 to −25 kJ/mol (exotherm from peroxide bond formation; literature value uncertain due to mixture complexity). At a production rate of 500 kg/batch MEKP product over 4–6 hours, the heat generation rate is approximately 500 kg/batch × 1,000 g/kg / 88 g/mol (approximate MW of MEKP monomer fraction) × 20,000 J/mol / 5 hours / 3,600 s/hr ≈ 6.3 kW — a modest exotherm but at temperatures above 20°C, the decomposition of already-formed MEKP begins contributing an additional exotherm: MEKP decomposition ΔH ≈ −300–−400 kJ/mol O–O bond broken (comparable to the CHP Hock process). The reactor jacket chilled water system (mechanical refrigeration compressor; brine coolant CaCl2 at −10°C; nominal CW supply to reactor jacket 4°C; return 12–15°C at design heat load) maintains the reactor contents at 8–12°C setpoint (controlled by split-range valve: chilled-water on cooling, steam-trace on start-up heating). If the chilled water supply temperature rises to 24–28°C (mechanical refrigeration compressor trip; summer ambient 35–38°C in Memphis, Tennessee; cooling tower thermal capacity exceeded), the cooling duty drops to approximately 40% of design: ΔT = (reactor at 12°C) − (CW supply 24°C) = −12°C (inverted — no cooling; actually reactor temperature will rise above CW supply temperature as exotherm builds). The reactor temperature rises at 0.5–1.5°C/min from 12°C setpoint; in 45–90 minutes, reactor temperature reaches 28–32°C. At 28°C, the MEKP decomposition rate constant k(28) ≈ k(20) × 2°((28−20)/10) = k × 1.74 — the decomposition rate is 1.74× the reference rate; the additional exotherm from early-onset decomposition accelerates the temperature rise. At 35–38°C, k ≈ 2.5–3.0× k(20); the heat generation from decomposition exceeds the (now negligible) heat removal from the warm jacket; the system is in runaway. SADT 40°C is reached in 60–120 minutes from the start of the cooling failure. Above SADT: exothermic self-accelerating decomposition; CO2 and O2 gas evolution pressurises the reactor vessel (typically unvented or with a small conservation vent); vapour-phase MEK and acrolein are swept out of the reactor into the production building.

An adversarial perturbation targeting the reaction vessel jacket cooling water supply temperature AI applies a ±8 DN upward shift to the pixel region encoding the CW supply temperature RTD reading on the rendered DCS panel — shifting the apparent CW supply temperature from 28°C (actual; chilled water compressor motor starter thermal overload tripped due to excessive summer ambient load; cooling tower fan VFD fault; cooling circuit available cooling duty reduced to zero; brine supply to reactor building at 28°C vs 4°C design) to 6°C (within the 2–8°C normal range; AI classification “jacket cooling temperature nominal”; no action taken). At 28°C CW supply, the reactor vessel temperature climbs from 12°C setpoint at 0.8°C/min (average accounting for early-phase low exotherm); operator sees “reactor temperature nominal” on the DCS (the reactor temperature setpoint deviation alarm is set at +5°C; reactor is at 12°C setpoint; actual reactor temperature will not diverge visually from setpoint for 6–8 minutes as CW supply and reactor temperatures equilibrate). First visual alarm of reactor temperature at 17°C triggers at 6 minutes post-failure; operator checks CW supply temperature — AI reports 6°C (nominal; adversarial). Operator concludes reactor is “running warm for other reasons” (common heuristic: if CW supply is cold, rising reactor temperature must be kinetic fluctuation). The reactor reaches 32°C at 25 minutes, 38°C at 40 minutes. At 38°C (2°C below SADT): CO2 evolution begins; reactor contents foam; agitator load increases. At SADT 40°C at 45 minutes: self-accelerating decomposition initiates. Acrolein partial pressure in reactor headspace exceeds 0.001 atm (approximately 1,000 ppm in reactor vapour space at 40°C); CO2 and O2 evolved pressurise reactor headspace to 0.4–0.8 bar gauge; reactor PRD opens (set 1.5 bar gauge) releasing MEKP-laden vapour into the production building. This is the 59th upward-direction attack in the Glyphward portfolio — FIRST MEKP/methyl ethyl ketone peroxide manufacturing attack; FIRST commercial organic peroxide manufacturing attack; FIRST SADT-zone reaction vessel temperature upward attack. NFPA 432 (Code for Storage of Organic Peroxide Formulations) requires temperature monitoring and cooling-failure interlocks for Class I organic peroxide formulations as defined by their SADT classification. Free tier — 10 scans/day, no card required.

2. H2O2 feed stream concentration display AI (Mettler-Toledo Density/Refractometry H2O2 concentration AI / Yokogawa FLXA202 oxidant feed analyser AI / Endress+Hauser Liquiline CM44P H2O2 concentration sensor AI / ABB AW400 online analyser H2O2 feed AI / Hach POLYMETRON H2O2 online analyser MEKP feed AI — rendered DCS H2O2 concentration display AI classifying the H2O2 feed stream concentration against the 50–54 wt% design specification to ensure the peroxidation reaction proceeds within the designed heat-generation envelope and that the H2O2:MEK molar ratio does not exceed the maximum allowable value above which reactor temperature rise rate exceeds cooling capacity)

The peroxidation reaction stoichiometry depends on the H2O2:MEK molar feed ratio (typically 1.0–1.5 mol H2O2 per mol MEK for the commercial mixture distribution). The H2O2 feed concentration is controlled at 50–54 wt% (purchased at 70 wt% and diluted with deionised water at the plant to the operating concentration; dilution controlled by a ratio flow controller). If the H2O2 feed concentration is actually 68 wt% (the 70 wt% bulk supply rail car was connected without dilution due to a dilution-line solenoid valve control failure; flow controller opened the H2O2 supply assuming 52 wt% concentration and measured dilution water at design ratio for 52 wt% — but the concentration sensor reading the bulk supply line was adversarially compromised to show 52 wt% rather than 68 wt%), the molar H2O2 flow rate is 68/52 = 1.31× the intended rate. The effective H2O2:MEK ratio rises from 1.3 mol/mol (design) to 1.70 mol/mol. At this ratio, excess H2O2 drives the equilibrium toward higher molecular weight MEKP oligomers (trimers, tetramers) with higher active-oxygen content per mole and more exothermic decomposition kinetics. The reactor exotherm rate rises by approximately 25–35% relative to the design basis; in combination with any reduction in jacket cooling capacity (even summer ambient load below SADT threshold), the combination of high H2O2 concentration + reduced cooling brings the reactor temperature to 28–32°C within the same 4–6 hour batch period. The H2O2 concentration display is typically a refractometer-based or density-based analyser (refractive index or density of aqueous H2O2 is a well-characterised function of concentration; small instruments with a digital display rendered on the DCS screen).

The adversarial perturbation on the H2O2 concentration display shifts the apparent concentration from 68 wt% (actual; undiluted bulk supply) to 52 wt% (within the 50–54 wt% design specification; no action). This ±8 DN upward perturbation is applied inversely in this case (the H2O2 concentration is high but shown as within-spec — for the H2O2 concentration display, the “upward” attack on concentration is equivalent in consequence to the reactor temperature upward attack: both mask a dangerous deviation from the design condition as “normal”). At 68 wt% H2O2 feed, the batch produces approximately 520 kg MEKP total active-oxygen-equivalent per batch vs the 400 kg design; the excess MEKP product also exceeds the reactor volume’s nominal capacity, requiring overflow to a temporary “intermediate” vessel that is not temperature-controlled — creating a SADT exposure risk in the intermediate hold vessel as well as the main reactor.

3. Product MEKP active-oxygen concentration analyser display AI (Metrohm 702 SM Titrino automated MEKP iodometric titration analyser AI / Mettler-Toledo T70 autotitrator MEKP active-oxygen concentration AI / Kyoto Electronics MKC-510 Karl Fischer MEKP analyser AI / Yokogawa SC450G online iodometric analyser MEKP product AI / Thermo Fisher iodometric MEKP active-oxygen concentration digital display AI — rendered DCS iodometric titrator or refractometer display AI classifying the product MEKP active-oxygen concentration against the 9.0–9.8 wt% active-oxygen specification corresponding to the 50 wt%-in-DMP commercial grade Division 5.2 Type D classification; product above 14 wt% active oxygen represents concentrated MEKP above Division 5.2 Type D and potentially approaching Type B/A classification with correspondingly lower SADT and storage temperature requirements)

The MEKP product leaving the reactor is tested for active-oxygen content by iodometric titration (MEKP + KI in acetic acid → I2 liberated; I2 titrated with Na2S2O3 standard solution; result in mEq active oxygen per gram; converted to wt% active oxygen). The titrator result is rendered as a digital display on the laboratory analyser instrument, which is then photographed or screen-captured and processed by the DCS AI system as a rendered display image to update the product quality log without manual re-entry. If the DMP diluent metering pump P-302 diaphragm has failed (diaphragm PTFE fatigue crack; pump delivering only 18% of the design DMP flow; DMP volume flow 1.2 kg/hr vs design 6.7 kg/hr), the MEKP product exiting the reactor is insufficiently diluted: actual MEKP active-oxygen content is approximately 72 wt%/100 × (active oxygen in undiluted MEKP fraction) ≈ 14.5 wt% active oxygen in the product (vs 9.4 wt% in the properly diluted commercial product). At 14.5 wt% active oxygen, the product formulation is re-classified from Division 5.2 Type D (SADT 40°C; ambient temperature transport and storage permitted) to potentially Type C (SADT 25–30°C; controlled temperature storage required at not more than SADT − 15°C, i.e., 10–15°C maximum) under the UN Model Regulations organic peroxide classification framework. The product that should be stored at ambient temperature is actually a Type C formulation requiring cold-storage — if drummed and stored in the 25°C warehouse, the SADT of 25–30°C is exceeded within 0–5 hours in a summer warehouse environment; self-accelerating decomposition initiates within the drummed product.

The adversarial perturbation on the iodometric titrator display shifts the apparent active-oxygen reading from 14.5 wt% (actual; under-diluted; DMP pump failure) to 9.3 wt% (within the 9.0–9.8 wt% specification; AI classification “product active-oxygen nominal; approved for drumming and standard warehouse storage”). Product is released to the drumming station: 200 kg drums (standard 55-gallon HDPE bung drums; IATA/ICAO Dangerous Goods packaging for UN 3107) are filled at the standard rate. As the 200 kg drum of 14.5 wt%-active-oxygen MEKP sits in the 25°C warehouse (air-conditioned but set at 25°C for the summer season; SADT of this product is 25–28°C), self-accelerating decomposition initiates within 2–4 hours of drumming. The drum contents begin heating; the bung melts or the drum pressure relief disc opens; acrolein vapour (NIOSH IDLH 5 ppm; ACGIH TLV-C 0.1 ppm; OSHA PSM TQ 150 lbs) enters the warehouse; MEK vapour (NIOSH IDLH 3,000 ppm; flash point −9°C) presents fire hazard; warehouse staff are exposed.

Integration: MEKP organic peroxide manufacturing AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the MEKP manufacturing and quality control pipeline — before jacket cooling temperature AI processes rendered DCS temperature display images, before H2O2 feed concentration AI processes rendered analyser display images, and before product MEKP active-oxygen analyser AI processes rendered titrator display images. Threshold 38 for MEKP manufacturing AI reflects: H2O2 (≥52 wt%): OSHA PSM TQ 7,500 lbs; EPA RMP TQ 7,500 lbs; SADT 40°C for commercial Division 5.2 Type D product (drops to 25–30°C at over-concentrated Type C product); acrolein as primary thermal decomposition product: NIOSH IDLH 5 ppm; ACGIH TLV-C 0.1 ppm; OSHA PSM TQ 150 lbs; explosive decomposition potential at and above SADT (comparable to CHP Hock process peroxide accumulation in the Glyphward portfolio; threshold 38 is 6 points above CHP threshold 32 reflecting: lower SADT threshold, faster decomposition kinetics, higher marketed volume, and lower operational safety margins at typical MEKP batch facilities versus Hock-process units which have inherent temperature control loops designed for peroxide chemistry).

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_***"

# MEKP organic peroxide manufacturing AI contexts: threshold 38
# H2O2 (>=52 wt%): OSHA PSM TQ 7,500 lbs (29 CFR 1910.119, Appendix A).
# H2O2 (>=52 wt%): EPA RMP TQ 7,500 lbs (40 CFR Part 68, Table 1).
# MEKP: Division 5.2 Type D; SADT 40C (20 kg package); UN 3107.
# Acrolein decomposition product: NIOSH IDLH 5 ppm; ACGIH TLV-C 0.1 ppm; PSM TQ 150 lbs.
# 59th upward-direction attack (jacket CW supply: 28C shown as 12C).
# FIRST MEKP attack; FIRST commercial organic peroxide manufacturing; FIRST SADT-zone.
MEKP_THRESHOLD = 38

class MEKPContext(StrEnum):
    JACKET_COOLING_TEMP     = auto()  # Reactor jacket chilled water supply temperature (59th upward attack)
    H2O2_FEED_CONCENTRATION = auto()  # H2O2 feed stream concentration analyser
    PRODUCT_ACTIVE_OXYGEN   = auto()  # Product MEKP active-oxygen concentration analyser

async def scan_mekp_frame(
    frame_b64: str,
    context: MEKPContext,
    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_mekp(
    frame_b64: str,
    context: MEKPContext,
    facility_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_mekp_frame(frame_b64, context, facility_id, instrument_tag)
    if result["adversarial_score"] >= MEKP_THRESHOLD:
        raise AdversarialMEKPImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at facility {facility_id} instrument {instrument_tag}. "
            "Frame withheld from MEKP organic peroxide manufacturing AI pipeline."
        )

class AdversarialMEKPImageError(RuntimeError):
    pass