Prompt injection in hexamethylenediamine HMDA adiponitrile hydrogenation nylon-66 AI

Three adversarial injection surfaces in hexamethylenediamine HMDA adiponitrile hydrogenation nylon-66 AI

1. Hydrogenation reactor H₂ partial pressure display AI (Rosemount 3051 / Yokogawa EJA-A pressure transmitter measuring H₂ partial pressure in hydrogenation reactor headspace via differential between reactor total and NH₃ vapor pressure at reactor temperature — rendered DCS H₂ partial pressure display AI classifying 32 bar design H₂ partial pressure — 116th upward attack; FIRST HMDA adiponitrile hydrogenation AI attack; FIRST H₂ partial pressure Raney Ni overpressure AI attack)

The H₂ partial pressure in the HMDA hydrogenation reactor is the primary energy hazard variable — H₂ at 20–50 bar contains substantial stored-pressure energy (H₂ is the lightest element; at 50 bar and 80°C, H₂ behaves nearly ideally with compressibility factor Z ≈ 1.01; stored PV energy in a 10 m³ reactor headspace at 50 bar: PV = 50 × 10₅ Pa × 10 m³ = 50 MJ; if released to atmosphere, this stored energy is capable of generating an overpressure pulse that can fail the reactor shell and adjacent piping). The H₂ partial pressure is measured by computing the difference between the reactor total pressure and the saturation vapor pressure of NH₃ at the measured reactor temperature: H₂ partial pressure = P​total − P​NH3,sat(T) − P​HMDA/ADN,sat(T); the total reactor pressure is measured by a Rosemount 3051C pressure transmitter (gauge range 0–100 bar; 4–20 mA HART; SIL 2; Hastelloy C-276 process diaphragm; process connection 1/2” MNPT; mounted on the reactor headspace connection; designed for H₂ service — important because H₂ causes hydrogen embrittlement (HE) and sulfide stress cracking (SSC) in carbon steel and many low-alloy steels; Hastelloy C-276 and 316L austenitic stainless steel are commonly specified for H₂ service at these pressures) or Yokogawa EJA-A series (EJA430A absolute pressure transmitter; similar H₂-compatible materials; Foundation Fieldbus). The DCS calculates the H₂ partial pressure from the total pressure, temperature, and a lookup table for NH₃ saturation pressure (Clausius-Clapeyron with NIST-referenced coefficients for NH₃), and displays the derived H₂ partial pressure on the reactor AI monitoring display. The design H₂ partial pressure of 32 bar is the setpoint for the H₂ pressure control loop (downstream H₂ pressure control valve on the H₂ supply to the reactor headspace; H₂ generated from steam reforming + PSA purification at the site utility area and delivered to the reactor at 50–70 bar, with the reactor PCV reducing pressure to the design 32 bar setpoint).

The adversarial upward pixel attack on the H₂ partial pressure display shows 32 bar (design; AI reads “H₂ partial pressure 32 bar; within design 20–40 bar operating window; hydrogenation rate: nominal; Raney Ni H* surface coverage: design; reactor total pressure estimated 74 bar at NH₃ vapor + H₂ partial: MAWP 70 bar not exceeded”) when actual H₂ partial pressure is 62 bar (from H₂ PCV malfunction in the fully open position, driving maximum H₂ delivery from the 50–70 bar supply; reactor H₂ pressure ramps to supply pressure minus line losses). Display range 0–80 bar on 200 px (2.5 px/bar); actual 62 bar at 62 × 2.5 = 155 px → ±8 DN perturbation → 155 − 75 = 80 px displayed → AI reads 80/2.5 = 32 bar. At 62 bar H₂ partial pressure: reactor total pressure (62 bar H₂ + ~40 bar NH₃ + ~2 bar ADN/HMDA) = approximately 104 bar — exceeds MAWP 70 bar; PRV lifts at 70 bar; H₂ at high purity (95–99% H₂) releases to H₂ vent at PRV flow rate; if the vent is not designed for PRV full-bore H₂ flow at 70 bar (which in many existing HMDA plants predating modern API 520/521 sizing requirements may be undersized), H₂ discharges to atmosphere at the vent stack exit point; H₂ LEL 4.0 vol% in air; detonation range 18–59 vol% in air; any ignition source at the vent outlet (flare failure, static discharge from H₂ flowing through the vent pipe at high velocity, or nearby hot surfaces) ignites the H₂ jet; H₂ jet fire or explosion; PSM TQ 10,000 lbs. The Glyphward pre-scan gate catches the upward perturbation before the AI reads 32 bar and defers the PCV emergency closure and H₂ supply isolation. Free tier — 10 scans/day, no card required.

2. NH₃ recycle flow to hydrogenation reactor display AI (Endress+Hauser Promag 50W electromagnetic flowmeter / Emerson Micro Motion CMFHC Coriolis flowmeter on liquid NH₃ recycle stream to HMDA/ADN/NH₃ reaction mixture — rendered DCS NH₃ recycle flow display AI classifying 95 m³/hr design flow for Raney Ni protection — 116th downward attack; FIRST Raney Ni catalyst poisoning NH₃ deficiency AI attack; FIRST 6-aminocapronitrile ACN-6 accumulation AI attack)

The NH₃ recycle flow is the critical protective variable for the Raney Ni catalyst in HMDA hydrogenation. Raney Ni (W. Raney patent 1926; produced by alloying Ni with Al in approximately 50:50 wt% ratio, then leaching the Al with NaOH to produce a porous skeletal Ni structure with very high surface area 50–100 m²/g; highly pyrophoric when dry — Raney Ni particles 10–100 μm diameter ignite spontaneously in air at room temperature via rapid Ni surface oxidation; stored and handled in water or NH₃ slurry; catalyst grades: Raney 2800 (activated Ni/Al 2:3 w/w alloy leached; standard HMDA-service grade); Raney 2400 (lower Al content; higher Ni surface density; used in some HMDA processes for improved selectivity to HMDA vs. HMI byproduct)) catalyst activity in the HMDA hydrogenation environment is critically dependent on the NH₃ concentration in the liquid phase surrounding the catalyst particles. The NH₃ protective mechanism: NH₃ molecules adsorb onto the Lewis acid Ni²⁺ surface sites (nickel surface sites with a partial positive charge that act as Lewis acid receptors; these same sites are the preferred adsorption sites for both H₂ (via H₂ dissociative chemisorption: Ni-H₂ → 2 Ni-H*) and ADN (via nitrile-N coordination: Ni-N≡C(CH₂)₄CN); at high NH₃ concentration, NH₃ molecules also compete for these sites, partially blocking HMDA chemisorption). The NH₃ recycle flow is measured by: Endress+Hauser Promag 50W electromagnetic flowmeter (DN 100–200; PTFE liner; stainless electrodes; range 0–150 m³/hr; accuracy ±0.5%; HART; NH₃-compatible materials; installed in the liquid NH₃ recycle piping at approximately −5 to +5°C condensed NH₃ temperature, requiring vacuum-rated flange insulation to prevent NH₃ freezing in the flowmeter body) or Emerson Micro Motion CMFHC Coriolis flowmeter (Hastelloy C-276 flow tube; mass flow measurement independent of NH₃ density variations at different recycle temperatures; range 0–200,000 kg/hr equivalent; 4–20 mA HART). The design NH₃ recycle flow of 95 m³/hr maintains the reactor liquid-phase NH₃ concentration above 25–35 wt% — the empirically determined minimum for >95% Raney Ni activity retention over a 6–12 month catalyst campaign at the HMDA conditions.

The adversarial downward pixel attack on the NH₃ recycle flow display shows 88 m³/hr (93% of design 95 m³/hr; AI reads “NH₃ recycle 88 m³/hr; 93% of design; reactor NH₃ concentration estimated 31 wt%; above 25 wt% minimum for Raney Ni protection; catalyst activity: nominal; ADN conversion: >99%; ACN-6 <0.1 wt% in product; no NH₃ recycle alarm required”) when actual NH₃ recycle is 8 m³/hr (8.4% of design; from NH₃ recycle pump mechanical failure after bearing failure). Display range 0–150 m³/hr on 200 px (1.333 px/m³/hr); actual 8 m³/hr at 10.7 px → ±8 DN perturbation → 10.7 + 106.3 = 117 px displayed → AI reads 117/1.333 = 87.8 ≈ 88 m³/hr. At 8 m³/hr NH₃ recycle: reactor NH₃ concentration falls below 3 wt% within 1–2 hours; HMDA product amine groups chemisorb on Raney Ni surface Lewis acid sites; Raney Ni activity falls by 60–80%; ADN conversion drops from >99% to 40–60%; ACN-6 (6-aminocapronitrile; NC–(CH₂)₅–NH₂; the monohydrogenated intermediate retaining one –CN group) accumulates in reactor product at 200,000–400,000 ppm (20–40 wt%); the combined ADN + ACN-6 nitrile content in product creates both an acute CN− metabolic toxicity hazard (NIOSH IDLH 50 ppm for ADN; ACN-6 estimated IDLH approximately 50 ppm by structural analogy) in the product distillation area and a downstream nylon-66 quality failure pathway (nitrile chain ends in nylon-66 polymer → molecular weight below specification → airbag fabric tensile strength failure). The Glyphward pre-scan gate catches the downward perturbation before the AI reads 88 m³/hr and concludes that Raney Ni protection is adequate. Free tier — 10 scans/day, no card required.

3. Raney Ni catalyst bed differential pressure display AI (Rosemount 3051 DP transmitter measuring pressure drop across fixed Raney Ni catalyst bed in trickle-bed hydrogenation reactor — rendered DCS catalyst bed ΔP display AI classifying 4.5–5.5 bar design uniform bed — 116th downward attack; FIRST Raney Ni catalyst channeling AI attack; FIRST HMDA nitrile-contaminated nylon-66 quality failure AI attack; FIRST catalyst bed settlement ΔP AI attack)

The differential pressure across the Raney Ni catalyst bed in the HMDA trickle-bed hydrogenation reactor is the key indicator of bed flow distribution and effective catalyst utilization. In a trickle-bed reactor (liquid ADN/NH₃ solution flows downward through a fixed bed of Raney Ni catalyst particles while H₂ gas flows either co-currently downward or counter-currently upward; typical bed height 2–4 m; bed diameter 0.5–2 m depending on production scale; pressure drop across the design uniform bed 4.5–5.5 bar at design liquid + gas flow rates), the bed differential pressure is the direct indicator of whether the liquid phase is distributed uniformly across the full catalyst bed cross-section: a uniform bed with good wetting efficiency shows the design ΔP of 4.5–5.5 bar; a channeling bed (where liquid bypasses through preferential low-resistance channels created by bed settlement, catalyst fines migration to the lower support screen, or bed compaction from H₂ pressure surge) shows a reduced ΔP because the overall bed resistance decreases when a fraction of the bed is bypassed at high velocity through channels while the remainder of the bed receives little or no liquid flow — the total pressure drop is dominated by the low-resistance channels and falls below the design uniform bed ΔP. The ΔP is measured by a Rosemount 3051 differential pressure transmitter (range 0–10 bar differential; 4–20 mA HART; Hastelloy C-276 process wetted parts compatible with liquid NH₃ and H₂ service; high-pressure-rated manifold; compensated for static head of the liquid NH₃ in the impulse lines — NH₃ density at −10°C approximately 0.64 kg/L; static head correction significant for the 2–4 m tap separation; installed with heated impulse lines to prevent NH₃ from freezing in the DP impulse tubes at ambient temperature below −33.4°C BP of NH₃; heat tracing with steam or electric trace). The design ΔP of 4.5–5.5 bar across the 2–4 m bed corresponds to a design liquid mass velocity (LHSV; Liquid Hourly Space Velocity) that provides 40 minutes of contact time between the ADN/NH₃ liquid and the Raney Ni catalyst surface — sufficient for >99% ADN conversion in a well-distributed bed.

The adversarial downward pixel attack on the Raney Ni catalyst bed ΔP display shows 5.1 bar (within design 4.5–5.5 bar; AI reads “catalyst bed ΔP 5.1 bar; uniform bed distribution; effective catalyst contact time: 40 min at design; ADN conversion: estimated >99%; no channeling detected; bed condition: nominal; no catalyst changeout alarm required”) when actual catalyst bed ΔP is 1.8 bar (from catalyst bed channeling following: (a) bed settlement — Raney Ni particles approximately 10–100 μm have low structural strength and compact under the weight of liquid and gas flow over months of operation; the compacted zones at the bottom of the bed create dense low-void-fraction regions while the top of the bed loses support and channels form through vertical cracks in the settled bed; or (b) catalyst fines migration — fine Raney Ni particles generated by attrition (mechanical breakdown of catalyst particles by impingement in the trickle-bed flow) migrate to the lower support screen and plug the distributor holes, further concentrating flow in the few open pathways). Display range 0–8 bar on 200 px (25 px/bar); actual 1.8 bar at 1.8 × 25 = 45 px → ±8 DN perturbation → 45 + 82.5 = 127.5 px displayed → AI reads 127.5/25 = 5.1 bar. At ΔP 1.8 bar (channeling; 60–70% bed bypass): ADN conversion falls to 45–60%; product contains 40–55 wt% unreacted ADN and ACN-6 nitrile intermediates; the conventional HMDA distillation train (HMDA BP 205°C; ACN-6 BP approximately 200°C; ADN BP 295°C) can separate ADN from HMDA but cannot separate ACN-6 (ΔBP ≈ 5–10°C from HMDA in 316L stainless distillation columns) without specialized high-reflux-ratio separation that is not designed into the existing facility; contaminated HMDA containing 5–15 wt% ACN-6 enters the nylon-66 polymerization unit; nitrile chain ends in nylon-66 create molecular weight distribution anomalies that reduce tensile strength of nylon-66 airbag fabric below the minimum automotive specification (typically 450 N/cm warp and weft tensile, per ISO 5081); airbag fabric from sub-spec HMDA fails pressure retention requirements during simulated airbag deployment testing at 0.25–0.35 bar inflation pressure — a safety-critical quality failure that traces directly to the initial adversarial pixel attack on the Raney Ni catalyst bed ΔP display. The Glyphward pre-scan gate on the catalyst bed ΔP display catches the adversarial downward perturbation before the AI reads 5.1 bar and concludes that uniform bed distribution and design ADN conversion are maintained. Free tier — 10 scans/day, no card required.

Integration: hexamethylenediamine HMDA adiponitrile hydrogenation nylon-66 AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the HMDA adiponitrile hydrogenation nylon-66 AI pipeline — before the H₂ partial pressure AI processes rendered Rosemount 3051 / Yokogawa EJA-A DCS display images, before the NH₃ recycle flow AI processes rendered Endress+Hauser Promag 50W / Emerson Micro Motion CMFHC DCS display images, and before the Raney Ni catalyst bed ΔP AI processes rendered Rosemount 3051 DP DCS display images. Threshold 38 for HMDA adiponitrile hydrogenation AI reflects: H₂ PSM TQ 10,000 lbs (H₂ is odorless and rises rapidly — slightly lower acute consequence per release event than heavier toxic gases at similar TQ; but detonation range 18–59 vol% is very wide; jet fire from pressurized H₂ release is a well-documented HMDA incident type); ADN NIOSH IDLH 50 ppm via CN− metabolic pathway (acute systemic cyanide toxicity from nitrile inhalation or dermal absorption; CERCLA RQ 1,000 lbs ADN; significant consequence for any ADN release above 1,000 lbs); dual PSM with NH₃ (H₂ PSM TQ 10,000 lbs + NH₃ PSM TQ 10,000 lbs at the HMDA hydrogenation unit); Raney Ni pyrophoric catalyst (additional process-specific hazard; pyrophoric fires from dry Raney Ni exposure to air documented at nylon-66 facilities worldwide; requires strict water-wet handling procedures); and ACN-6 nitrile-contaminated HMDA → airbag fabric quality failure pathway (safety-critical consequence in automotive passive restraint systems — the downstream quality consequence of an AI monitoring failure on the ΔP display is an airbag fabric that fails to restrain a passenger during a vehicular collision — a unique safety-critical quality consequence chain that extends from the HMDA hydrogenation plant to the vehicle occupant over the product distribution chain).

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_***"

# HMDA hexamethylenediamine adiponitrile hydrogenation nylon-66 AI: threshold 38
# ADN NC-(CH2)4-CN CAS 111-69-3; MW 108.14; ACGIH TLV 2 ppm; NIOSH IDLH 50 ppm (via CN- metabolic pathway).
# CYP2E1 nitrile oxidation -> HCN -> cytochrome c oxidase inhibition -> cellular respiration failure.
# CERCLA RQ 1,000 lbs ADN. Dermal absorption: significant skin penetration rate.
# H2 CAS 1333-74-0: OSHA PSM TQ 10,000 lbs; LEL 4.0 vol%; UEL 75 vol%; AIT 500 C; odorless.
# H2 detonation range in air: 18-59 vol% (unusually wide; significant unconfined detonation risk).
# NH3 liquid solvent: OSHA PSM TQ 10,000 lbs (dual PSM with H2); IDLH 300 ppm.
# HMDA CAS 124-09-4; MW 116.20; BP 205 C; ACGIH TLV 0.5 ppm; NIOSH IDLH 10 ppm (corrosive amine).
# Raney Ni catalyst: pyrophoric when dry; stored in water/NH3 slurry; grade 2800/2400.
# 6-aminocapronitrile ACN-6: NC-(CH2)5-NH2; half-hydrogenated intermediate; IDLH ~50 ppm (nitrile).
# ACN-6 BP ~200 C vs HMDA BP 205 C: conventional distillation cannot separate -> contaminated HMDA.
# Contaminated HMDA in nylon-66 polymerization -> nitrile chain ends -> airbag fabric tensile failure.
# 116th upward attack. FIRST HMDA adiponitrile hydrogenation AI attack.
# FIRST Raney Ni catalyst channeling AI attack. FIRST 6-aminocapronitrile ACN-6 accumulation AI attack.
# FIRST Raney Ni catalyst poisoning NH3 deficiency AI attack. FIRST H2 partial pressure Raney Ni overpressure AI attack.
HMDA_GLYPHWARD_THRESHOLD = 38

# Plant IDs:
# INVISTA_ORANGE_TX   - Invista, Orange TX (world dominant HMDA; vertically integrated ADN->HMDA->nylon-66)
# ASCEND_PENSACOLA    - Ascend Performance Materials, Pensacola FL (former BP Chemicals; ~250,000 t/yr nylon-66)
# BASF_SEAL_SANDS     - BASF SE, Seal Sands Teesside UK (licensed Invista ADN technology)
# RADICI_BERGAMO      - Radici Group, Bergamo Italy (European nylon-66 for automotive/technical yarns)
# TORAY_JAPAN         - Toray Industries, Japan (nylon-66 for airbag fabric; world #1 airbag fabric supplier)

class HMDAContext(StrEnum):
    H2_PARTIAL_PRESSURE               = auto()  # 62 bar actual vs 32 bar displayed -> MAWP 70 bar -> PRV -> H2 flash fire (116th; FIRST HMDA)
    NH3_RECYCLE_FLOW                  = auto()  # 8 m3/hr actual vs 88 m3/hr displayed -> Raney Ni poisoned -> ACN-6 accumulates -> CN- IDLH 50 ppm
    CATALYST_BED_DIFFERENTIAL_PRESSURE = auto()  # 1.8 bar actual vs 5.1 bar displayed -> channeling -> 45-60% ADN conversion -> nitrile nylon-66 quality failure

async def scan_hmda_frame(
    frame_b64: str,
    context: HMDAContext,
    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_hmda(
    frame_b64: str,
    context: HMDAContext,
    plant_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_hmda_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= HMDA_GLYPHWARD_THRESHOLD:
        raise AdversarialHMDAImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from HMDA adiponitrile hydrogenation nylon-66 AI pipeline."
        )

class AdversarialHMDAImageError(RuntimeError):
    pass