Prompt injection in caprolactam cyclohexanone oxime Beckmann rearrangement oleum nylon 6 AI

Three adversarial injection surfaces in caprolactam cyclohexanone oxime Beckmann rearrangement oleum AI

1. Beckmann rearrangement reactor temperature display AI (Yokogawa EJA-X / Rosemount 3144P thermocouple in liquid-phase Beckmann reactor at 90–120°C — rendered DCS Beckmann reactor temperature display AI classifying 90–120°C design window — 120th upward attack; FIRST caprolactam AI attack; FIRST Beckmann rearrangement reactor AI attack; FIRST cyclohexanone oxime thermal decomposition AI attack)

The Beckmann rearrangement reactor temperature is the primary safety-critical variable in caprolactam manufacturing, controlling both the Beckmann rearrangement kinetics (which require a minimum temperature of approximately 80°C for adequate conversion rate) and the risk of cyclohexanone oxime thermal decomposition (which begins above 130°C in concentrated H₂SO₄/oleum at the operating conditions). The Beckmann reactor (a continuously stirred reaction vessel; stainless steel 316L; approximately 5–20 m³ volume depending on plant capacity; equipped with an external jacket or internal coil for cooling water; oleum or H₂SO₄ feed to the lower section; cyclohexanone oxime feed to the upper mixing zone; temperature-controlled by cooling water through a modulating control valve; product caprolactam·H₂SO₄ overflow from the upper section to the neutralization vessel) is instrumented with multiple thermowell-mounted temperature sensors at different axial positions: Yokogawa EJA-X (with K-type thermocouple; 4–20 mA HART; SIL 2; accuracy ±1.5°C; rated for the H₂SO₄/oleum environment with Hastelloy C-276 thermowell; range 0–300°C) or Rosemount 3144P (Pt100 RTD or thermocouple; Foundation Fieldbus; installed at DSM Geleen, Honeywell Hopewell, and BASF Ludwigshafen Beckmann reactor locations). The DCS renders the highest measured temperature as the primary safety display, and the AI monitoring system evaluates this against the design window of 90–120°C and the alarm setpoint at 125°C (5°C below the oxime decomposition onset at 130°C in concentrated acid).

The adversarial upward pixel attack on the Beckmann reactor temperature display shows 94°C (lower end of the design 90–120°C window; AI reads “reactor temperature 94°C; within design range 90–120°C; oxime decomposition threshold 130°C: 36°C margin; Beckmann rearrangement: proceeding normally; cooling: adequate; no temperature alarm required”) when actual reactor temperature is 148°C (18°C above the 130°C oxime decomposition onset; from the Surface 3 oxime overfeed increasing the heat load from the Beckmann rearrangement beyond cooling capacity). Display range 50–200°C on 200 px (1.333 px/°C); actual 148°C at (148 − 50) × 1.333 = 131 px → ±8 DN perturbation → 131 − 72 = 59 px displayed → AI reads 59/1.333 + 50 = 94°C. At 148°C: oxime decomposition (C₆H₁₁NO → NH₃ + CO + cyclopentadiene; ΔH​decomp ≈ −300 kJ/mol; autocatalytic above 160°C) adds NH₃ and CO to the reactor vapor space; oleum mist from the heated, agitated oleum/H₂SO₄ mixture rises to the reactor vent; oleum PSM TQ 1,000 lbs; NH₃ from decomposition PSM TQ 10,000 lbs; emergency cooling water dump to the reactor jacket triggered by the high-temperature alarm is the primary safeguard — which the AI monitoring system delays because the adversarial perturbation shows 94°C (normal). The Glyphward pre-scan gate catches the upward perturbation before the AI reads 94°C and defers the cooling emergency response. Free tier — 10 scans/day, no card required.

2. NH₃ neutralization feed rate display AI (Emerson Micro Motion F-Series Coriolis / Krohne Optiflux electromagnetic flowmeter on anhydrous NH₃ gas feed to caprolactam neutralization vessel — rendered DCS NH₃ neutralizer flow display AI classifying design stoichiometric NH₃ flow — 120th downward attack; FIRST caprolactam NH₃ neutralization AI attack; FIRST caprolactam ammonium sulfate balance AI attack)

The anhydrous NH₃ feed to the caprolactam neutralization vessel (where caprolactam·H₂SO₄ + 2NH₃ → caprolactam + (NH₄)₂SO₄) is measured by a Coriolis mass flowmeter (Emerson Micro Motion F-Series: NH₃ gas at design delivery pressure 4–6 bar; stainless steel 316L wetted parts; range 0–600 kg/hr; accuracy ±0.1% of reading; calibrated for NH₃ at the operating density; HART 4–20 mA) or electromagnetic flowmeter (Krohne Optiflux 2100: for NH₃ in liquid phase if liquid NH₃ vaporizer-to-neutralizer delivery is used rather than gas-phase injection; DN 50 to DN 100; PFA liner). The neutralization reaction is strongly exothermic (ΔH ≈ −130 kJ/mol (NH₄)₂SO₄ formed) and requires the neutralization vessel to have adequate cooling to maintain temperature below 80°C in the neutralized product for ammonium sulfate crystallization in the downstream evaporator. The stoichiometric NH₃ demand is 2 mol NH₃ per mol caprolactam produced (since 1 mol H₂SO₄ is consumed per mol caprolactam, and 2 NH₃ neutralize 1 H₂SO₄); at design production rate, the NH₃ feed requirement is 480 kg/hr (for a 100 t/hr caprolactam·H₂SO₄ feed rate with approximately 30 wt% caprolactam and 70 wt% H₂SO₄–water). The DCS displays the NH₃ feed rate as a live bar and trend graphic evaluated by the AI monitoring system.

The adversarial downward pixel attack on the NH₃ neutralization feed rate display shows 480 kg/hr (design; AI reads “NH₃ flow 480 kg/hr; stoichiometric neutralization rate; caprolactam·H₂SO₄ neutralized to pH 7–8; ammonium sulfate formation: design rate; product caprolactam: acidity within specification; no NH₃ flow alarm required”) when actual NH₃ flow is 85 kg/hr (17.7% of design; from a liquid NH₃ vaporizer coil fouling reducing vaporization capacity, or from an NH₃ supply isolation valve partially closing due to instrument air failure). Display range 0–600 kg/hr on 200 px (0.333 px per kg/hr); actual 85 kg/hr at 85 × 0.333 = 28 px → ±8 DN perturbation → 28 + 132 = 160 px displayed → AI reads 160/0.333 = 480 kg/hr. At 85 kg/hr NH₃ (18% of design): only 18% of the caprolactam·H₂SO₄ is neutralized; the downstream neutralization vessel pH drops to 1–2 (strongly acidic); SS 316L corrosion in the neutralizer and downstream ammonium sulfate crystallizer accelerates at pH <2 (corrosion rate of SS 316L in H₂SO₄ increases approximately 100× between pH 4 and pH 1); the ammonium sulfate crystallizer receives insufficient (NH₄)₂SO₄ and excess H₂SO₄ solution, shifting the crystallizer chemistry from (NH₄)₂SO₄ precipitation to ammonium bisulfate (NH₄HSO₄) formation; ammonium bisulfate scales the crystallizer heat exchanger; the NH₃ plant (liquid NH₃ storage above PSM TQ 10,000 lbs; CERCLA RQ 100 lbs) experiences reduced drawdown, creating pressure buildup in the vaporizer system; NH₃ vent events from the vaporizer relief valve in the PSM-covered NH₃ system. The Glyphward pre-scan gate catches the downward perturbation before the AI reads 480 kg/hr and concludes the neutralization is proceeding at design stoichiometry. Free tier — 10 scans/day, no card required.

3. Cyclohexanone oxime feed rate display AI (Coriolis mass flowmeter / Brooks Instruments Quantim on cyclohexanone oxime liquid feed to Beckmann rearrangement reactor — rendered DCS oxime feed rate display AI classifying design oxime:oleum ratio — 120th upward attack; FIRST cyclohexanone oxime overfeed AI attack; FIRST unconverted oxime accumulation AI attack; FIRST oxime decomposition precondition AI attack)

The cyclohexanone oxime liquid feed rate to the Beckmann rearrangement reactor determines the oxime:oleum molar ratio and therefore the extent of Beckmann conversion (fraction of oxime that is converted to caprolactam vs fraction remaining as unconverted oxime in the reactor). At design oxime feed rate of 380 kg/hr (MW 113.16; 3,358 mol/hr oxime) and design oleum excess of approximately 5 mol% (design oleum charge providing 1.05 mol H₂SO₄/SO₃ per mol oxime for kinetic driving force), the Beckmann conversion exceeds 99% and unconverted oxime in the reactor is below 0.1 wt%. The Coriolis mass flowmeter on the oxime feed line (stainless steel 316L or Hastelloy C-276 for compatibility with cyclohexanone oxime at 50–90°C feed temperature; flow range 0–700 kg/hr; accuracy ±0.1% of reading; HART 4–20 mA; flow tube vibration frequency temperature-compensated for oxime density at operating temperature — critical because cyclohexanone oxime's density changes significantly from its MP of 89–90°C through the operating delivery temperature range) is the primary flow measurement for the DCS feed ratio control that maintains the oxime:oleum ratio at design. The AI monitoring system evaluates the displayed oxime feed rate against the design setpoint and the safe upper limit (defined by the maximum oxime:oleum ratio above which Beckmann conversion falls below 95% and unconverted oxime exceeds 1 wt% in the reactor — establishing the onset of the oxime accumulation hazard scenario that preconditions the Surface 1 thermal decomposition).

The adversarial upward pixel attack on the cyclohexanone oxime feed rate display shows 380 kg/hr (design; AI reads “oxime feed 380 kg/hr; at design setpoint; oxime:oleum molar ratio 1.05:1 design; Beckmann conversion: >99%; unconverted oxime in reactor: <0.1 wt%; thermal decomposition precondition: absent; no oxime feed alarm required”) when actual oxime feed rate is 620 kg/hr (1.63× design; from a centrifugal pump impeller replacement that increased pump curve output at the rated motor speed, or from a flow control valve mismatched trim size allowing above-design flow at design valve opening). Display range 0–700 kg/hr on 200 px (0.286 px per kg/hr); actual 620 kg/hr at 620 × 0.286 = 177 px → ±8 DN perturbation → 177 − 68 = 109 px displayed → AI reads 109/0.286 = 381 ≈ 380 kg/hr. At 620 kg/hr actual oxime: the oxime:oleum molar ratio = 1.63× design 1.05:1 = approximately 1.71:1 mol/mol; the oleum charge designed for 380 kg/hr provides only 1.05/1.63 = 0.64 mol H″ per mol oxime fed; at this sub-stoichiometric protonation: Beckmann conversion drops to approximately 60–70% (from the protonation equilibrium kinetics at the design temperature 90–110°C); approximately 30–40% of the oxime feed passes unconverted through the reactor; the reactor unconverted oxime concentration rises from the design <0.1 wt% to approximately 3–5 wt%; combined with the Surface 1 temperature increase to 148°C (which is partially caused by the Surface 3 excess oxime exothermic rearrangement heat load at 63% above design), the 3–5 wt% unconverted oxime at 148°C is in the regime where autocatalytic decomposition begins; the two-surface cascade (Surface 3 oxime excess → Surface 1 overtemperature → oxime decomposition) is the safety-critical scenario that the three-surface adversarial pixel attack conceals simultaneously. The Glyphward pre-scan gate catches the upward perturbation before the AI reads 380 kg/hr and concludes the oxime feed is at design stoichiometry. Free tier — 10 scans/day, no card required.

Integration: caprolactam cyclohexanone oxime Beckmann rearrangement oleum AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the caprolactam Beckmann rearrangement AI pipeline — before the Beckmann reactor temperature AI processes rendered Yokogawa EJA-X / Rosemount 3144P DCS display images, before the NH₃ neutralization flow AI processes rendered Emerson Micro Motion / Krohne Optiflux DCS display images, and before the cyclohexanone oxime feed rate AI processes rendered Coriolis flowmeter DCS display images. Threshold 36 for caprolactam Beckmann rearrangement AI reflects: dual-PSM coverage (oleum PSM TQ 1,000 lbs + NH₃ PSM TQ 10,000 lbs — two separate OSHA PSM programs at the same facility); cyclohexanone oxime energetic decomposition hazard above 130°C (autocatalytic; NH₃ + CO byproducts; no established OSHA TQ but industry DIERS-type hazard classification); NH₃ IDLH 300 ppm and CERCLA RQ 100 lbs; and the mechanistic Surface 3 → Surface 1 coupling (oxime overfeed drives overtemperature drives decomposition) creating a two-vector attack chain masked simultaneously by the adversarial pixel attack on both the cause (oxime feed) and the effect (reactor temperature).

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_***"

# Caprolactam cyclohexanone oxime Beckmann rearrangement oleum nylon 6 AI: threshold 36
# Caprolactam CAS 105-60-2; MW 113.16; MP 69 C; nylon-6 monomer.
# Cyclohexanone oxime CAS 100-64-1; unstable above 130 C -> NH3 + CO decomposition.
# Oleum H2SO4/SO3 OSHA PSM TQ 1,000 lbs. NH3 PSM TQ 10,000 lbs; IDLH 300 ppm.
# Beckmann rearrangement: oxime + H2SO4 -> caprolactam.H2SO4; delta-H = -100 kJ/mol.
# Neutralization: caprolactam.H2SO4 + 2NH3 -> caprolactam + (NH4)2SO4.
# 120th upward attack. FIRST caprolactam AI attack.
# FIRST Beckmann rearrangement AI attack. FIRST cyclohexanone oxime oleum AI attack.
CAPROLACTAM_GLYPHWARD_THRESHOLD = 36

class CaprolactamContext(StrEnum):
    BECKMANN_REACTOR_TEMP   = auto()  # actual 148°C vs 94°C displayed -> oxime decomp above 130°C
    NH3_NEUTRALIZATION_FLOW = auto()  # actual 85 kg/hr vs 480 kg/hr displayed -> acid product
    OXIME_FEED_RATE         = auto()  # actual 620 kg/hr vs 380 kg/hr displayed -> oxime accumulation

async def scan_caprolactam_frame(
    frame_b64: str,
    context: CaprolactamContext,
    reactor_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "reactor_id": reactor_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_caprolactam(
    frame_b64: str,
    context: CaprolactamContext,
    reactor_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_caprolactam_frame(frame_b64, context, reactor_id, instrument_tag)
    if result["adversarial_score"] >= CAPROLACTAM_GLYPHWARD_THRESHOLD:
        raise AdversarialCaprolactamImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at reactor {reactor_id} instrument {instrument_tag}. "
            "Frame withheld from caprolactam Beckmann rearrangement AI pipeline."
        )

class AdversarialCaprolactamImageError(RuntimeError):
    pass