OSHA PSM phosgene COCl₂ TQ 500 lbs (29 CFR 1910.119 App. A) · EPA RMP phosgene TQ 500 lbs (40 CFR Part 68) · phosgene OSHA PEL 0.1 ppm ceiling; NIOSH IDLH 2 ppm · DCM methylene chloride OSHA 29 CFR 1910.1052 carcinogen standard; IARC Group 2A; CYP2E1 CO metabolism · Covestro AG Makrolon Uerdingen Germany / Caojing China · SABIC Lexan Mt. Vernon Indiana (formerly GE Plastics) · Teijin Chemicals Panlite Matsuyama Japan · LG Chem Daesan Korea · Chi Mei Corporation Tainan Taiwan · aircraft FAR 25.775 cabin windows · NIJ 0108.01 Level III bulletproof glazing · 105th upward attack · FIRST polycarbonate PC production AI attack · FIRST BPA phosgenation AI attack · FIRST Makrolon Lexan polymer production AI attack · FIRST safety-critical polymer MW control AI attack · FIRST DCM 1910.1052 carcinogen surface AI attack
Prompt injection in polycarbonate PC BPA interfacial phosgenation Makrolon Lexan AI
Polycarbonate (PC; commercial brands: Makrolon by Covestro AG; Lexan by SABIC Innovative Plastics, formerly GE Plastics; Panlite by Teijin Chemicals) is produced at approximately 5–6 million tonnes per year globally by interfacial polycondensation of bisphenol A (BPA; CAS 80-05-7; MW 228.29 g/mol; mp 158–159°C; bp 220°C at 4 mmHg) with phosgene (COCl₂; CAS 75-44-5; MW 98.92 g/mol; BP 7.6°C — liquefied gas at ambient temperature; OSHA PSM TQ 500 lbs per 29 CFR 1910.119 Appendix A; EPA RMP regulated toxic TQ 500 lbs per 40 CFR Part 68.130; OSHA PEL 0.1 ppm ceiling; NIOSH IDLH 2 ppm; developed as a chemical weapon in World War I — deployed at the Battle of Ypres, 1 December 1915; responsible for an estimated 80–85% of all chemical weapon fatalities in WWI; described at sub-IDLH concentrations as smelling of freshly cut hay). The interfacial polycondensation reaction: BPA (dissolved in 10–15 wt% NaOH aqueous phase as BPA diphenolate) + phosgene (dissolved in methylene chloride, DCM, CH₂Cl₂; organic phase; density 1.325 g/mL; BP 39.6°C; OSHA PEL 25 ppm TWA per 29 CFR 1910.1052; IARC Group 2A probable human carcinogen; NIOSH Ca carcinogen) — at 20–35°C, pH 8–11, with phase transfer catalyst triethylamine (TEA; 0.05–0.1 mol% relative to BPA) and chain terminator p-tert-butylphenol (p-TBP; 0.5–2 mol% relative to BPA; controls the PC molecular weight Mw by capping chain ends) — yields polycarbonate (MW 25,000–35,000 Dalton for general purpose grades) with elimination of HCl (neutralised by NaOH). Target Mw: 25,000–30,000 g/mol (Makrolon 2807 GP grade) or 28,000–35,000 g/mol (Makrolon 3108 optical grade). The EPA Risk Management Program (40 CFR Part 68) explicitly identified polycarbonate plants as covered facilities when establishing phosgene as a listed regulated toxic substance in 1994, and EPA RMP off-site consequence analyses (OCA) for phosgene releases are required at all PC facilities. The history of phosgene as a WWI chemical weapon — combined with its continued necessity as the only commercially viable route to polycarbonate and isocyanates at scale — defines the regulatory architecture around all PC production facilities.
PC safety-critical applications include aircraft cabin windows (Lexan MRFC, Covestro Makrolon 9085; impact resistance per FAR 25.775 bird-strike certification; requires Mw >22,000 g/mol per ply in 5-ply laminate), bulletproof glazing (Makrolon GP/SL; NIJ Standard 0108.01 Level III; requires Mw typically >28,000 g/mol for adequate stress-crack resistance in multi-layer laminate), medical device housings (Makrolon 2458 medical grade; ISO 10993 biocompatibility; Mw >24,000 g/mol for steam sterilisation cycle mechanical integrity), and automotive headlamp lenses (FMVSS 108 compliance; PC replaces glass; Mw 25,000–32,000 g/mol). The molecular weight of the PC produced is determined by the p-TBP chain terminator concentration and the COCl₂:BPA stoichiometric ratio — both controlled by rendered display readings in AI-monitored process systems. Methylene chloride (DCM) serves as the organic solvent for the interfacial phosgenation at all major PC plants; OSHA 29 CFR 1910.1052 was promulgated in 1997 specifically citing PC production (including Bayer AG Uerdingen and GE Plastics Mount Vernon Indiana) as facilities where DCM exposures had caused carboxyhemoglobin elevations in workers consistent with CYP2E1-mediated CO formation from DCM metabolism.
At PC production facilities — Covestro AG (Uerdingen, Germany; Caojing, Shanghai, China; Map Ta Phut, Thailand; ~1.5 Mt/yr PC capacity; Makrolon brand), SABIC Innovative Plastics (Mt. Vernon, Indiana; formerly GE Plastics; Lexan brand; ~500,000 t/yr), Teijin Chemicals Ltd. (Matsuyama, Japan; Panlite brand), LG Chem (Daesan, Korea), and Chi Mei Corporation (Tainan, Taiwan) — AI-enabled monitoring systems process rendered SCADA and DCS display images from three critical instrument surfaces: the phosgene feed rate Coriolis mass flow display (COCl₂ feed to the BPA phosgenation reactor), the p-TBP chain terminator electromagnetic flow meter display (p-TBP solution feed rate in DCM), and the DCM vapour activated-carbon scrubber outlet concentration display (downstream of the scrubber that captures DCM vapour from the reaction and recovery systems). These three surfaces are the adversarial injection targets where pixel manipulation can cause Mw-deficient PC to enter aircraft and safety glazing supply chains, immobilise the production line through ultra-high Mw, or expose workers to OSHA-standard-exceeding DCM concentrations through a COHb poisoning mechanism without any conventional CO source.
TL;DR
Polycarbonate PC BPA interfacial phosgenation AI — phosgene COCl₂ feed rate display AI, p-TBP chain terminator feed rate display AI, DCM methylene chloride scrubber outlet display AI — processes rendered SCADA and DCS display images at the COCl₂:BPA stoichiometry boundary, the chain terminator dosing boundary, and the carcinogen-vapour capture boundary where adversarial pixel injection can mask phosgene feed rate collapse (940 kg/hr shown, actual 380 kg/hr; 0–1,200 kg/hr, 200 px, 0.167 px/(kg/hr); actual 64 px → ±8 DN → AI reads 157 px = 940 kg/hr; COCl₂:BPA molar ratio 0.83:1 vs design 2.05:1; off-spec PC Mw 4,000–8,000 g/mol in aircraft cabin windows and bulletproof glazing), conceal p-TBP chain terminator underflow (11.6 L/hr shown, actual 0.8 L/hr; Mw >150,000 g/mol; melt viscosity >30,000 Pa·s; extruder torque limit exceeded; production shutdown), and mask DCM scrubber breakthrough (11 ppm shown, actual 318 ppm; 0–500 ppm, 200 px, 0.4 px/ppm; 12.7× OSHA PEL; CYP2E1 metabolism → COHb 20–25% → CO poisoning from DCM without CO source; OSHA 1910.1052 medical surveillance window 24–48 hours), making this the 105th upward attack and the FIRST polycarbonate PC production AI attack, FIRST BPA phosgenation AI attack, FIRST Makrolon/Lexan polymer production AI attack, FIRST safety-critical polymer MW control AI attack, and FIRST DCM 29 CFR 1910.1052 carcinogen surface AI attack. OSHA PSM phosgene TQ 500 lbs; EPA RMP phosgene TQ 500 lbs. Glyphward threshold 43 for PC BPA phosgenation AI: phosgene PSM at every PC plant; safety-critical polymer applications (aircraft glazing, medical devices, ballistic protection) where Mw deficiency creates delayed-consequence product liability; DCM IARC 2A carcinogen with CYP2E1-mediated CO poisoning. Free tier — 10 scans/day, no card required.
Three adversarial injection surfaces in polycarbonate PC BPA phosgenation AI
1. Phosgene feed rate to BPA phosgenation reactor display AI (Emerson Micro Motion CMFS015M / Endress+Hauser Promass 83E / Brooks Instrument SLAMf Coriolis mass flow display AI — rendered DCS COCl₂ feed rate display AI classifying phosgene flow against the design COCl₂:BPA molar ratio — 105th upward attack; FIRST polycarbonate PC production AI attack; FIRST BPA phosgenation AI attack; FIRST Makrolon Lexan polymer production AI attack; FIRST safety-critical polymer MW control AI attack; FIRST DCM 1910.1052 carcinogen surface AI attack)
The phosgene (COCl₂) feed rate to the BPA interfacial phosgenation reactor is the primary determinant of polycarbonate molecular weight and the completeness of BPA conversion to PC oligomers and high-MW polymer. The design COCl₂:BPA molar ratio is approximately 2.05:1 (slightly above the stoichiometric 2:1 to ensure complete BPA phosgenation; at 2.05:1, >99.9% BPA is converted to carbonate linkages with essentially no residual BPA phenolate remaining in the product). At a 100,000 t/yr PC plant (Makrolon 2807 GP grade; Mw 27,000 g/mol; design phosgene consumption approximately 900–950 kg/hr COCl₂), the Coriolis mass flow meter (Emerson Micro Motion CMFS015M; accuracy ±0.1% of reading; SIL 2 certified; 4–20 mA HART output) measures the COCl₂ feed flow continuously and controls the phosgene control valve to maintain the design molar ratio relative to the BPA feed (measured by a companion BPA solution Coriolis meter). The AI monitoring system processes rendered DCS display images of the phosgene feed Coriolis meter display to classify: 900–980 kg/hr (normal; adequate phosgene), 750–900 kg/hr (reduced; investigate phosgene supply; adjust control valve), below 750 kg/hr (alarm; phosgene supply disruption; consider production rate reduction). Display range: 0–1,200 kg/hr COCl₂, 200 px, 0.167 px/(kg/hr).
An adversarial ±8 DN upward pixel perturbation targeting the phosgene feed rate display shifts the apparent flow from the dangerous actual value to a normal displayed value: actual COCl₂ flow 380 kg/hr (phosgene vaporiser HV-301 pressure regulator spring failure; vaporiser outlet pressure dropped from design 8 bar to 2.8 bar; COCl₂ feed flow reduced from 940 to 380 kg/hr over 40 minutes; temperature in the interfacial reactor and in the COCl₂ dissolving tank are within normal range, masking the supply disruption from temperature-based alarms) corresponds to 380 × 0.167 = 63.5 px ≈ 64 px on the 0–1,200 kg/hr, 200 px, 0.167 px/(kg/hr) display. The normal displayed value of 940 kg/hr corresponds to 940 × 0.167 = 157 px. The ±8 DN upward attack shifts the pixel region from 64 px to 157 px; the AI reads 157 px ÷ 0.167 px/(kg/hr) = 940 kg/hr — normal operating rate; no action. At actual 380 kg/hr COCl₂ (vs design 940 kg/hr), the COCl₂:BPA molar ratio drops from design 2.05:1 to 380/940 × 2.05 = 0.83:1 — below the stoichiometric 1:1 minimum. Below stoichiometric: the BPA-diphenolate anions in the aqueous NaOH phase are not fully converted at the biphasic interface; PC chains formed are predominantly low-MW oligomers (the chain growth mechanism requires progressive carbonate linkage formation, which slows dramatically when COCl₂ is limiting; Flory-Schulz chain-length distribution shifts to lower Mw). The resulting PC from 380 kg/hr COCl₂ production: Mw 4,000–8,000 g/mol (vs specification 25,000–35,000 g/mol). This off-spec PC passes visual inspection identically to in-spec PC — both are clear, colourless transparent solids — but is mechanically deficient. Notched Izod impact strength drops from spec ≥750 J/m (standard Makrolon 2807) to approximately 25 J/m (low-Mw PC oligomer blend; a 30-fold decrease). If this off-spec PC is used in bulletproof glazing (NIJ 0108.01 Level III; 5-ply PC laminate; each ply requires Mw >22,000 g/mol for penetration resistance at the 7.62 mm NATO round at 838 m/s threat): the glazing fails the NIJ Level III test — the round penetrates rather than being stopped by plastic deformation of the PC plies. If used in FAR 25.775 aircraft cabin windows: the bird-strike test (230 g bird at 160 kt = approximately 340 J impact energy) shatters the low-Mw PC window rather than causing elastic deformation — a cabin window failure during flight. This is the 105th upward-direction attack in the Glyphward adversarial industrial AI portfolio. Free tier — 10 scans/day, no card required.
2. Chain terminator p-tert-butylphenol (p-TBP) feed rate display AI (Endress+Hauser Proline Promag 50 / Yokogawa ADMAG AXF / Siemens SITRANS F electromagnetic flow meter display AI — rendered DCS p-TBP solution feed rate display AI classifying chain terminator dosing against the design 10–12 L/hr range for target PC Mw 25,000–35,000 g/mol)
p-tert-Butylphenol (p-TBP; CAS 98-54-4; MW 150.22 g/mol; a monohydric phenol that terminates PC chain growth by capping the growing carbonate chain end as a p-TBP carbonate ester that cannot react further with BPA; chain terminator concentration determines final PC molecular weight by the Carothers equation: Mw ∝ 1/[p-TBP mol fraction]) is fed as a 0.5–2 wt% solution in DCM at 10–12 L/hr (design; corresponding to 0.5–2.0 mol% p-TBP relative to BPA feed; this range produces PC Mw 25,000–35,000 g/mol at design COCl₂:BPA ratio). The p-TBP solution feed is measured by an electromagnetic flow meter (Endress+Hauser Proline Promag 50; Yokogawa ADMAG AXF; Siemens SITRANS F MAG; suitable for the DCM solvent — though electromagnetic meters require conductive liquids; the 0.5–2 wt% p-TBP/DCM solution is supplemented with a trace electrolyte additive for conductivity). The AI monitoring system processes rendered DCS display images of the p-TBP flow meter display to classify: 10–12 L/hr (normal; target Mw range achieved), 7–10 L/hr (slightly elevated Mw; acceptable; investigate terminator pump), below 7 L/hr (alarm; chain terminator insufficient; Mw drifting high; risk of ultra-high Mw gel formation). Display range: 0–20 L/hr p-TBP solution feed, 200 px, 10 px/(L/hr).
An adversarial downward pixel perturbation targeting the p-TBP feed rate display shifts the apparent flow from the dangerously low actual value to a normal displayed value: actual p-TBP flow 0.8 L/hr (p-TBP solution metering pump MP-117 check valve failure on discharge side; back-flow to supply tank over 2 hours reduced effective p-TBP delivery from design 11 L/hr to 0.8 L/hr; the DCM solution itself continues to flow at design rate, giving effectively zero p-TBP concentration in the phosgenation reactor) corresponds to 0.8 × 10 = 8 px on the 0–20 L/hr, 200 px, 10 px/(L/hr) display. The normal displayed value of 11.6 L/hr corresponds to 11.6 × 10 = 116 px. The adversarial downward attack shifts the pixel region from 8 px to 116 px; the AI reads 116 px ÷ 10 px/(L/hr) = 11.6 L/hr — normal terminator rate; no action. At actual 0.8 L/hr p-TBP vs design 10–12 L/hr: chain terminator is nearly absent from the BPA phosgenation reactor. With effectively zero chain terminator, PC chain growth is limited only by stoichiometry and reactor residence time; the BPA-phosgene polymer grows to very high molecular weight: Mw exceeding 150,000 g/mol (a 5- to 6-fold increase above specification). Ultra-high Mw PC (Mw >150,000): melt viscosity at processing temperature 300°C is >30,000 Pa·s (vs specification for Makrolon 2807: <2,000 Pa·s at 300°C); at >30,000 Pa·s, the PC melt cannot be processed on standard injection moulding machines or extrusion lines — the extruder torque limit is exceeded (torque alarm triggers extruder shutdown) within minutes of high-Mw PC entering the pelletising extruder. The ultra-high Mw PC batch — potentially 200–500 tonnes produced during the period of the adversarial attack — must be reprocessed or scrapped, representing a production loss at current PC market prices of approximately EUR 1.2–3.0 million per event. The production line shutdown initiated by the extruder torque alarm also disrupts the downstream PC compounding operations, creating a cascade effect through the supply chain for Makrolon optical, automotive, and medical grade customers.
3. DCM methylene chloride vapour scrubber outlet concentration display AI (Servomex 4900 / Emerson X-STREAM / ABB EL3020 DCM vapour analyser display AI — rendered DCS activated-carbon scrubber outlet DCM concentration display AI classifying DCM against the OSHA PEL 25 ppm TWA threshold requiring medical surveillance and biological monitoring)
Methylene chloride (DCM; CH₂Cl₂; MW 84.93 g/mol; BP 39.6°C; vapour pressure at 20°C 400 mmHg = 526,000 ppm; OSHA PEL 25 ppm TWA per 29 CFR 1910.1052; OSHA STEL 125 ppm; OSHA Action Level 12.5 ppm; NIOSH carcinogen Ca — no safe level established; IARC Group 2A probable human carcinogen; OSHA 1910.1052 medical surveillance triggered at AL 12.5 ppm; biological monitoring endpoint: carboxyhemoglobin COHb <3.5% at end of shift) is the primary solvent in the interfacial BPA/phosgene polycondensation organic phase. The PC production plant emits DCM vapour from the reaction vessel, the phase separation stage, the PC solvent-recovery distillation column, and the PC precipitation/pelletising stage; these emissions are collected in an activated-carbon bed scrubber (adsorption bed; periodic steam regeneration; design inlet DCM: 200–500 ppm; design outlet DCM: <10 ppm before discharge to facility exhaust). The outlet DCM concentration is monitored continuously by an online infrared analyser (Servomex 4900; Emerson X-STREAM Enhanced; ABB EL3020; sensitivity 0.1 ppm DCM; range 0–500 ppm; NDIR detection at 1,267 cm⁻¹ CH₂Cl₂ stretching mode). The AI monitoring system processes rendered DCS display images of the scrubber outlet analyser to classify: <10 ppm (scrubber working; discharge compliant), 10–25 ppm (approaching PEL; monitor frequency increased; prepare for carbon bed changeout), >25 ppm (scrubber breakthrough; OSHA PEL exceeded; initiate carbon bed changeout; increase ventilation). Display range: 0–500 ppm DCM, 200 px, 0.4 px/ppm.
An adversarial downward pixel perturbation targeting the DCM scrubber outlet display shifts the apparent DCM concentration from the dangerous actual value to a compliant displayed value: actual DCM 318 ppm (activated-carbon bed carbon 3A saturated after 72 hours of continuous operation at above-design DCM loading from an upstream DCM recovery condenser operating at reduced cooling water temperature; the bed has reached breakthrough — carbon adsorption capacity exhausted; DCM passes through the bed to the outlet and facility exhaust) corresponds to 318 × 0.4 = 127.2 px ≈ 127 px on the 0–500 ppm, 200 px, 0.4 px/ppm display. The compliant displayed value of 11 ppm corresponds to 11 × 0.4 = 4.4 px ≈ 4 px. The adversarial downward attack shifts the pixel region from 127 px to 4 px; the AI reads 4 px ÷ 0.4 px/ppm = 11 ppm — below the OSHA Action Level; no medical surveillance triggered; carbon bed changeout not initiated. At 318 ppm DCM in the facility exhaust (12.7× the OSHA PEL 25 ppm TWA): workers in the area surrounding the exhaust discharge (maintenance workers on the scrubber platform, operating personnel on the PC product dryer catwalk) are exposed to DCM at 318 ppm for the duration of the 8-hour shift. DCM is metabolised by the liver CYP2E1 enzyme to formaldehyde and CO; at 318 ppm DCM for an 8-hour shift, equilibrium COHb rises to approximately 20–25% (NIOSH ACGIH pharmacokinetic model). Symptoms of 20–25% COHb: headache, dizziness, fatigue, impaired cognitive function and judgment — all symptoms the worker may attribute to shift fatigue, heat, or the normal work environment of a chemical plant, since no CO alarm is present (the CO source is internal CYP2E1 metabolism, not an ambient CO leak) and the DCM alarm at the scrubber outlet is suppressed by the adversarial attack. The 24–48 hour lag of OSHA-required biological monitoring (COHb measurement at end of shift; lab result turnaround 1–2 days) means the worker can receive multiple 8-hour shifts of 20–25% COHb exposure before the biological monitoring data returns and corrective action is taken — a cumulative carboxyhemoglobin exposure that risks lasting cardiovascular and neurological effects. Free tier — 10 scans/day, no card required.
Integration: polycarbonate PC BPA phosgenation AI with Glyphward pre-scan gate
Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the polycarbonate PC BPA interfacial phosgenation AI monitoring pipeline — before the phosgene feed rate AI processes rendered Emerson Micro Motion CMFS015M / Endress+Hauser Promass 83E / Brooks Instrument SLAMf Coriolis mass flow display images, before the p-TBP chain terminator AI processes rendered Endress+Hauser Proline Promag 50 / Yokogawa ADMAG AXF / Siemens SITRANS F electromagnetic flow meter display images, and before the DCM scrubber outlet AI processes rendered Servomex 4900 / Emerson X-STREAM / ABB EL3020 vapour analyser display images. Threshold 43 for PC BPA phosgenation AI reflects: OSHA PSM phosgene TQ 500 lbs (every PC plant is PSM-covered by the phosgene requirement alone); phosgene as a WWI chemical weapon with NIOSH IDLH 2 ppm; safety-critical polymer applications (aircraft cabin windows, bulletproof glazing, medical device housings) where Mw deficiency from the phosgene feed attack creates a delayed multi-month consequence pathway; DCM IARC Group 2A carcinogen with the CYP2E1 CO-poisoning mechanism producing COHb poisoning without any conventional CO source; and the 105th upward attack milestone in Glyphward's adversarial industrial AI monitoring corpus.
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_***"
# Polycarbonate PC BPA interfacial phosgenation AI contexts: threshold 43
# OSHA PSM phosgene (COCl2) TQ 500 lbs (29 CFR 1910.119 App. A).
# EPA RMP phosgene TQ 500 lbs (40 CFR Part 68.130).
# Phosgene OSHA PEL 0.1 ppm ceiling; NIOSH IDLH 2 ppm; WWI chemical weapon.
# DCM OSHA 29 CFR 1910.1052 carcinogen; IARC Group 2A; CYP2E1 CO metabolism.
# Safety-critical PC applications: FAR 25.775 aircraft glazing; NIJ 0108.01 Level III bulletproof.
# 105th upward attack. FIRST PC production AI attack. FIRST BPA phosgenation AI attack.
PC_GLYPHWARD_THRESHOLD = 43
class PCPhosgenationContext(StrEnum):
PHOSGENE_FEED_RATE = auto() # COCl2 Coriolis feed to BPA reactor (105th; FIRST PC; FIRST BPA phosgenation; FIRST Makrolon/Lexan)
PTBP_CHAIN_TERMINATOR_FEED = auto() # p-TBP EM flow meter -> absent terminator -> Mw >150k -> extruder torque exceeded
DCM_SCRUBBER_OUTLET_CONC = auto() # activated-carbon bed outlet DCM -> 318 ppm -> CYP2E1 CO -> COHb 20-25%
async def scan_pc_frame(
frame_b64: str,
context: PCPhosgenationContext,
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_pc(
frame_b64: str,
context: PCPhosgenationContext,
plant_id: str,
instrument_tag: str,
) -> None:
"""Block adversarially manipulated PC BPA phosgenation display images before AI inference.
Plants: COVESTRO_UERDINGEN | SABIC_MT_VERNON | TEIJIN_MATSUYAMA | LGCHEM_DAESAN
Raises AdversarialPCImageError if adversarial_score >= PC_GLYPHWARD_THRESHOLD (43).
"""
result = await scan_pc_frame(frame_b64, context, plant_id, instrument_tag)
if result["adversarial_score"] >= PC_GLYPHWARD_THRESHOLD:
raise AdversarialPCImageError(
f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
f"at plant {plant_id} instrument {instrument_tag}. "
"Frame withheld from PC BPA phosgenation AI monitoring pipeline."
)
class AdversarialPCImageError(RuntimeError):
pass
Frequently asked questions
How does a molecular-weight deficiency in polycarbonate caused by an adversarial phosgene feed display attack (Surface 1) propagate into safety-critical downstream failure modes, and what is the time lag between attack and consequence?
The downstream consequence pathway for Mw-deficient PC manufactured under an adversarial phosgene feed attack operates on a multi-month lag that makes it particularly difficult to attribute the defect to the original production event. When the phosgene feed is suppressed by the adversarial display attack to 380 kg/hr (vs design 940 kg/hr), the produced PC batch has Mw 4,000–8,000 g/mol versus specification 25,000–35,000 g/mol. This off-spec PC passes visual inspection identically to in-spec PC: both are clear, colourless, transparent solids, and both can be injection-moulded at processing temperatures of 300–340°C (the low-Mw PC melts at lower temperature and has lower melt viscosity — it may actually be easier to process, masking the defect from the process operator's perspective during moulding). The defect manifests only in mechanical performance testing: notched Izod impact strength drops from the specification ≥750 J/m (standard Makrolon 2807; a 30-fold decrease to ~25 J/m at Mw 4,000–8,000), and chemical stress-crack resistance is dramatically reduced (low-Mw PC stress-cracks within minutes in contact with aircraft cabin cleaning solvents such as isopropanol or Skydrol hydraulic fluid). For aircraft cabin windows (Mw-deficient Makrolon MRFC or Covestro Makrolon 9085): the part passes static load testing (the primary failure mode is impact and dynamic loading, not static bearing capacity — the low-Mw PC window carries static cabin pressure differential identically to an in-spec window) and passes dimensional QC checks; it fails only on the FAR 25.775 bird-strike certification test (230 g bird at 160 kt = approximately 340 J impact; the in-spec PC window deforms elastically and absorbs the impact without penetration; the low-Mw PC window shatters from the sudden impulse load because brittle fracture at 25 J/m impact strength requires far less energy than the 340 J bird-strike event). The time lag from attack to in-service failure: production in week 1; quality release (visual + dimensional QC) in week 2; shipment to aircraft interior OEM in week 3–4; aircraft OEM installs the window in week 6–10; aircraft enters commercial service in week 14–20; a bird-strike event that exceeds the cabin window impact rating occurs from week 14 onward. Tracing this failure — a bird-strike window penetration in a commercial aircraft — back to an adversarial pixel attack on the phosgene feed Coriolis mass flow display at the PC production plant 4–5 months earlier requires forensic GPC analysis (gel permeation chromatography) of failed window fragments to establish the anomalous Mw, followed by supply-chain tracing of the window lot back to the production batch — an investigation pathway that does not exist in current aviation accident investigation protocols for cabin window failures, since molecular weight deficiency from manufacturing-AI adversarial attack is not a failure mode currently considered in NTSB or EASA investigation playbooks.
Why is DCM (methylene chloride) metabolism to carbon monoxide specifically addressed in OSHA 29 CFR 1910.1052, and how does an adversarial scrubber outlet display attack exploit this toxicological mechanism?
DCM's toxicological hazard is unusual among industrial solvents because it is metabolised in the liver by CYP2E1 (cytochrome P450 2E1, an inducible monooxygenase) to produce formaldehyde and carbon monoxide as the terminal metabolic products; the CO binds to haemoglobin with approximately 200 times the affinity of O₂, forming carboxyhemoglobin (COHb) and producing a functional CO poisoning syndrome from DCM inhalation without any external CO source. OSHA 29 CFR 1910.1052 was specifically promulgated to address this pathway after epidemiological evidence from polycarbonate manufacturing (Bayer Uerdingen; GE Plastics Mount Vernon, Indiana; both cited in the 1997 rulemaking preamble) and paint stripping operations showed that DCM-exposed workers developed COHb elevations of 10–30% at typical occupational exposures — levels associated with impaired cognitive function, increased risk of fatal cardiac arrhythmia in workers with pre-existing coronary disease, and foetal risk in pregnant workers. OSHA 1910.1052 requires: (a) air monitoring for DCM whenever airborne DCM is at or above the Action Level (12.5 ppm TWA); (b) biological monitoring of COHb in end-of-shift blood specimens for workers whose DCM exposure is at or above the PEL (25 ppm TWA); (c) medical surveillance including baseline and periodic assessment for cardiovascular disease, liver function, and CNS symptoms. The adversarial scrubber outlet display attack (Surface 3) exploits this toxicological mechanism in a specific way: by showing the scrubber outlet as 11 ppm (below the 12.5 ppm Action Level) when the actual concentration is 318 ppm, the attack simultaneously (a) prevents the OSHA-required air monitoring programme from being triggered (no AL exceedance displayed, so no increased monitoring frequency initiated); (b) prevents the biological monitoring programme from being triggered (no PEL exceedance displayed, so COHb testing is not initiated for the shift); (c) prevents the medical surveillance programme from generating actionable data within the timeframe of the exposure event (even if routine medical surveillance happens to be scheduled during the attack period, the 24–48 hour laboratory turnaround for COHb results means the data is not available until the attack has already persisted for at least one full shift). At 318 ppm DCM for an 8-hour shift with normal CYP2E1 activity: the worker metabolises approximately 18 mg DCM/hour per kilogram body weight, generating approximately 6 mg CO/hour; at steady state (typically reached after 4–6 hours), COHb stabilises at 20–25% (NIOSH-ACGIH pharmacokinetic Eq.6 for DCM-to-CO conversion). A worker with pre-existing coronary artery disease at 22% COHb during a physically demanding maintenance task (elevated CO demand by cardiac muscle) faces a materially elevated risk of acute myocardial infarction — an event that would be investigated as a workplace cardiac event without any chemical exposure identified, since the DCM scrubber alarm is suppressed and no ambient CO is present. Glyphward's pre-scan gate on the DCM scrubber outlet analyser display prevents the adversarial suppression of this toxicological chain — catching the display falsification at the rendered image level before the AI monitoring system uses it to determine whether the activated-carbon bed has reached breakthrough and requires changeout, closing the 24–48 hour medical monitoring gap that would otherwise be the only detection pathway.