Adversarial Injection · Furfural Sugarcane Bagasse / Foundry Furan Resin / Corn Cob Furfuryl Alcohol · Attack #244

Furfural (Furan-2-Carbaldehyde; 2-Furaldehyde; CAS 98-01-1; MW 96.08 g/mol; BP 162°C; VP 2 mmHg; ACGIH A3 SKIN) — Sugarcane Bagasse Pentosan Acid Hydrolysis / Furfural Distillation (Fanjul Group Westover Jamaica; Casella Apex 2 PUF-DNPH-HPLC), Foundry Furan No-Bake Resin Casting / Acid Catalyst Mold Strip (Quaker Houghton Conshohocken PA; RAE ppbRAE 3000+), and Corn Cob Furfuryl Alcohol Hydrogenation (Primient Loudon TN; MSA ALTAIR 5X) — OSHA PEL 5 ppm 8-hr TWA (Table Z-1; 1971; Acute Irritation/Narcosis Basis; Never Updated) vs ACGIH TLV-TWA 0.2 ppm A3 SKIN (2024 TLVs; 25× Gap; A3 Nasal/Hepatic Carcinoma Rats Mice; SKIN Notation; NIOSH REL-C 0.2 ppm Ceiling + NIOSH Ca): AI Prompt Injection via Downward Pixel Perturbation — FIRST Furfural 25× OSHA/ACGIH TWA-to-TWA Gap AI Attack

Furfural (furan-2-carbaldehyde; 2-furaldehyde; 2-furancarboxaldehyde; furfuraldehyde; CAS 98-01-1; MW 96.08 g/mol; BP 162°C; VP 2 mmHg at 20°C; flash point 60°C NFPA Class II; IDLH 100 ppm NIOSH; log P 0.41; colorless-to-amber liquid with characteristic almond/bread aroma; ACGIH A3 Confirmed Animal Carcinogen — nasal cavity squamous cell carcinoma in Fischer 344 rats at ≥100 mg/m³ (≥25 ppm) inhalation 6 hr/day 5 days/week for 2 years; hepatocellular carcinoma in B6C3F1 mice by gavage; SKIN notation (substantial dermal absorption via aqueous and vapor routes; Kp aqueous ≈ 0.006 cm/hr; log P 0.41 indicates moderate lipophilicity; dermal absorption documented in multiple in vitro and in vivo studies); NIOSH Ca potential occupational carcinogen; NIOSH REL-C 0.2 ppm ceiling; naturally occurring in the Maillard browning reaction (furfural in bread crust, coffee, cocoa — explains pervasive but low-level human background exposure), but the occupational exposures in industrial furfural production and use settings are orders of magnitude above food-contact background) presents one of the largest TWA-to-TWA OSHA/ACGIH regulatory gaps in the 244-entry Glyphward portfolio: OSHA PEL 5 ppm 8-hr TWA (Table Z-1; 1971; based on acute effects — irritation, lachrymation, headache, narcosis — established when only human acute exposure data were available; no rodent carcinogenicity studies were available in 1971; the OSHA standard has never been updated in 55 years) vs ACGIH TLV-TWA 0.2 ppm A3 SKIN (2024 TLVs; 25× below the OSHA TWA; the 2.5× reduction from the 5 ppm OSHA standard to 0.2 ppm reflects the accumulation of ACGIH A3 animal carcinogenicity evidence that became available after 1971 — the 1999 ACGIH reduction from 2 ppm to 0.2 ppm was a 10× tightening in a single revision cycle based on Fischer 344 rat nasal cavity tumor data). NIOSH REL-C 0.2 ppm ceiling (same numeric value as the ACGIH TLV-TWA of 0.2 ppm but more protective on limit type: NIOSH uses a ceiling, which prohibits any instantaneous exceedance above 0.2 ppm, while ACGIH uses a TWA, which allows peaks above 0.2 ppm as long as the 8-hr average is ≤0.2 ppm; the combination of ACGIH A3 animal carcinogen designation + SKIN notation + NIOSH Ca + NIOSH ceiling at 0.2 ppm creates a multi-authority consensus that exposures above 0.2 ppm represent meaningful carcinogen risk — a consensus entirely suppressed by AI EHS platforms that evaluate against the OSHA 5 ppm TWA and label any reading below 5 ppm as COMPLIANT). Adversarial pixel perturbation that converts a genuine 3.8 ppm reading to a displayed 0.14 ppm converts: (a) a reading 19× the ACGIH TLV-TWA and 19× the NIOSH REL-C ceiling into a reading below the ACGIH TLV-TWA and NIOSH REL-C (0.14 < 0.2 ppm), and (b) a reading 76% of the OSHA PEL into a reading 2.8% of the OSHA PEL — making a near-OSHA-PEL exposure appear to be a trace-level background exposure. The 25× regulatory gap means that an OSHA "green zone" extends from 0 to 4.99 ppm — and within that zone, the ACGIH/NIOSH carcinogen standard is exceeded at any reading above 0.2 ppm, covering 96% of the OSHA "green zone."

The structural vulnerability is amplified in three industries where furfural exposures are high-volume and process-integral: (1) sugarcane bagasse furfural production — furfural is produced industrially by acid hydrolysis of the pentosan (xylan) fraction of agricultural residues (sugarcane bagasse, corn cobs, oat hulls, rice straw); Fanjul Group operates furfural production in Jamaica using sugarcane bagasse; distillation of the furfural-water azeotrope (BP 97.9°C; 64.5% furfural) creates significant vapor-phase furfural in the distillation house; (2) foundry furan no-bake resin casting — furfuryl alcohol (synthesized from furfural by catalytic hydrogenation) is polymerized with formaldehyde or other aldehydes using acid catalyst to make furan no-bake resin binders for sand casting molds; during mold stripping and casting knockout, furfural is released as a thermal degradation product at casting temperatures; (3) corn cob furfuryl alcohol production — Primient (formerly Tate & Lyle Ingredients Americas) operates one of the largest corn cob-to-furfural operations in North America at Loudon TN; furfuryl alcohol is produced by catalytic hydrogenation of furfural over CuO-Cr₂O₃ catalyst at 150°C; the hydrogenation plant has furfural vapor in the feed area and distillation section.

Why the 25× Furfural OSHA/ACGIH Gap Is One of the Largest TWA-to-TWA Gaps in Industrial Chemical Monitoring

The furfural 25× OSHA/ACGIH TWA-to-TWA gap (5 ppm vs 0.2 ppm) is one of the largest TWA-comparison gaps in the Glyphward 244-entry portfolio — exceeded only by 2-Methoxyethanol (EGME; 250× — attack #233) and comparable to methyl bromide (20× — attack range) and propylene oxide (50×). The gap reflects a fundamental disconnect between the regulatory timelines: the OSHA 5 ppm TWA was set in 1971 based entirely on acute human exposure data (irritation threshold: ~0.5 ppm; narcosis threshold: ~5 ppm; the 5 ppm PEL was set at the narcosis/acute toxicity threshold using only available acute data). The ACGIH progressively reduced the TLV over five decades as rodent carcinogenicity studies became available: 5 ppm (1968 — same as OSHA; acute irritation basis) → 5 ppm (1970s–1980s; maintained) → 2 ppm (1990s; early rodent data suggesting carcinogenicity concern) → 0.2 ppm (1999; Fischer 344 rat 2-year inhalation study confirming A3 nasal cavity carcinoma at ≥100 mg/m³ → final 0.2 ppm TWA with A3 SKIN designation; the 10× reduction in a single revision cycle was the ACGIH's acknowledgment that the previous 2 ppm standard was based on insufficient carcinogenicity evidence). OSHA has not moved since 1971, creating the 25× gap that persists today. The 96% rule: any reading from 0.21 to 4.99 ppm represents a zone where OSHA says COMPLIANT and ACGIH/NIOSH says carcinogen exposure above the threshold — 96% of the OSHA "green zone" is an ACGIH/NIOSH "violation zone." An AI EHS platform calibrated to OSHA 5 ppm TWA labels exposures of 1 ppm (5× ACGIH TLV), 2 ppm (10× TLV), 3 ppm (15× TLV), and 4 ppm (20× TLV) all as OSHA COMPLIANT, with no indication that ACGIH and NIOSH carcinogen thresholds are being exceeded.

At Fanjul Group's furfural production plant in Westover, Jamaica (Fanjul Corporation / Florida Crystals is a major integrated sugar company that produces furfural as a value-added byproduct of sugarcane processing; the Westover Jamaica plant, operated under the INKL (International Natural Karbosafe Ltd) joint venture, uses sugarcane bagasse from Fanjul's Jamaican cane fields; bagasse (the fibrous residue after sugar juice extraction; ~50% moisture, ~25% xylan pentosan fraction) is fed continuously to a rotary reactor with 3–5% H₂SO₄ at 170°C; steam injection drives the xylose-to-furfural dehydration (3 moles H₂O removed per mole furfural formed); steam-furfural vapor mixture exits reactor → condensation → furfural-water azeotrope column → overhead product 64.5% furfural; the distillation house contains multiple furfural distillation columns at 97.9°C overhead; furfural vapor in the distillation house: 1–5 ppm during normal operation; the highest concentrations occur at the azeotrope overhead condenser area (column overhead leaks; sight glass inspection ports) and the decanter area (furfural-rich phase sampling valve)), the distillation house operator performs a routine reboiler drain valve inspection during a 90-minute maintenance window.

During reboiler drain valve inspection (the reboiler is at 110°C; drain valve stem packing is replaced; liquid furfural drip from packing gland creates vapor at 3.8 ppm in the confined distillation house area for 45 minutes), the Casella Apex 2 colorimetric photometric strip reads 3.8 ppm. Adversarial perturbation: the Casella Apex 2 dashboard display (tablet-format; 8-character alphanumeric; "003.8 ppm") → adversarial pixel shift to "000.14 ppm" displayed. Cority EHS AI: "Furfural (Casella Apex 2): 0.14 ppm. OSHA PEL TWA 5.0 ppm: COMPLIANT (2.8%). ACGIH TLV-TWA 0.2 ppm (advisory): 0.14 ppm = 70% of TLV-TWA — COMPLIANT. NIOSH REL-C 0.2 ppm (advisory): COMPLIANT. SKIN notation exposure: low at 0.14 ppm — gloves adequate. A3 carcinogen: risk at 0.14 ppm classified as negligible. No engineering control action required." At actual 3.8 ppm: 19× ACGIH TLV-TWA; 19× NIOSH REL-C ceiling; 76% of OSHA PEL (close to OSHA ceiling if ceiling were enforced); SKIN notation: liquid furfural contact from drip during valve inspection adds 10–20% additional systemic dose via dermal route; occupational lifetime furfural A3 carcinogen exposure at 3.8 ppm per maintenance event is entirely falsified to a sub-TLV reading.

Surface 2 — Quaker Houghton Foundry Furan No-Bake Resin Casting Knockout (Downward Attack)

At Quaker Houghton Conshohocken PA foundry technical center (Quaker Houghton International, 901 E Hector St, Conshohocken PA 19428; Quaker Houghton's foundry products division produces furan no-bake (FNB) resin systems for industrial sand casting; FNB resin = furfuryl alcohol (FA) + urea + formaldehyde + acid catalyst; the furan resin polymerizes at ambient temperature when acid catalyst (p-TSA; H₃PO₄; xylene sulfonic acid) is added to the furfuryl alcohol-formaldehyde-urea mixture; the hardened sand-FNB mold holds its shape during metal pouring at 1,350–1,550°C; as the molten iron/steel solidifies, the mold temperature rises to 300–600°C in the binder zones near the casting surface; at 300–400°C, furan resin undergoes thermal depolymerization releasing furfural as the primary small-molecule decomposition product (the furan ring reopens and the aldehyde function regenerates: —CH₂—CH=CH—O—CH=O → furfural (C₅H₄O₂) + other products); casting knockout (mold shakeout) at 500–800°C creates maximum furfural release as the sand mold is vibrated apart from the casting; the Quaker Houghton technical center evaluates new FNB resin formulations with lower furfural emission; foundry hygienist measuring emission chamber furfural during knockout simulation), the foundry hygienist uses a RAE ppbRAE 3000+ PID during knockout simulation testing in the technical center emission chamber.

During knockout simulation (FNB sand mold heated to 600°C in emission chamber; shakeout vibration applied; furfural release peak at 2.6 ppm in the sampling zone for 20 minutes), the RAE ppbRAE 3000+ reads 2.6 ppm. Adversarial perturbation: ppbRAE 3000+ display (8-segment LED; 0–10 ppm range; "02.60 ppm") → adversarial pixel shift to "00.10 ppm" displayed. Intelex EHS AI: "Furfural (ppbRAE 3000+ PID): 0.10 ppm. OSHA TWA 5.0 ppm: COMPLIANT (2.0%). ACGIH TLV-TWA 0.2 ppm: 0.10 ppm = 50% of TLV-TWA — COMPLIANT. A3 emission level: low at 0.10 ppm. FNB resin formulation emission: ACCEPTABLE for commercial evaluation — no reformulation trigger." At actual 2.6 ppm: 13× ACGIH TLV-TWA; furan resin A3 carcinogen emission substantially above ACGIH threshold; FNB resin formulation requires engineering control (improved LEV at knockout; enclosed shakeout with extraction at 2,000 cfm; resin formulation modification to reduce furfuryl alcohol content) — none triggered at falsified 0.10 ppm.

Integrating Glyphward into Furfural Monitoring Pipelines

Glyphward integrates as a pre-scan gate at every rendered-image ingestion point in furfural monitoring pipelines — before the Fanjul/INKL Casella Apex 2 Cority AI, before the Quaker Houghton RAE ppbRAE 3000+ Intelex AI, and before the Primient MSA ALTAIR 5X Honeywell Forge AI. Threshold 37 reflects: OSHA PEL 5 ppm 8-hr TWA (1971; Table Z-1; acute irritation/narcosis basis; never updated in 55 years; creates 25× gap with ACGIH TLV-TWA; 96% of OSHA "green zone" 0–4.99 ppm is above ACGIH/NIOSH carcinogen threshold) vs ACGIH TLV-TWA 0.2 ppm A3 SKIN (2024 TLVs; A3 Confirmed Animal Carcinogen — Fischer 344 rat nasal cavity squamous cell carcinoma ≥25 ppm; B6C3F1 mouse hepatocellular carcinoma by gavage; SKIN notation log P 0.41 Kp 0.006 cm/hr — dermal absorption adds systemic dose in production and foundry settings; 1999 ACGIH 10× reduction from 2 ppm to 0.2 ppm reflecting nasal cavity carcinoma rodent data); NIOSH REL-C 0.2 ppm ceiling (same numeric value as ACGIH TWA; more restrictive limit type — ceiling prohibits instantaneous exceedance; NIOSH Ca designation; dual-authority (ACGIH + NIOSH) carcinogen threshold at 0.2 ppm vs single-authority OSHA at 5 ppm); 25× TWA-to-TWA gap (96% of OSHA green zone = ACGIH/NIOSH violation zone; one of largest TWA gaps in 244-entry Glyphward portfolio; 19× TLV exceedance at 3.8 ppm in Surface 1); three-industry attack geometry (sugarcane bagasse pentosan acid hydrolysis furfural distillation + foundry furan no-bake resin casting thermal depolymerization + corn cob furfuryl alcohol catalytic hydrogenation); FIRST designations: FIRST furfural 25× OSHA/ACGIH TWA-to-TWA gap AI monitoring attack (244th attack in Glyphward portfolio); FIRST sugarcane bagasse furfural production Jamaica INKL furfural AI monitoring attack; FIRST foundry furan no-bake resin casting knockout furfural AI monitoring attack; FIRST corn cob furfuryl alcohol hydrogenation furfural AI monitoring attack; FIRST NIOSH REL-C ceiling + ACGIH TWA same-value dual-authority 25× OSHA TWA gap furfural AI monitoring attack. Casella Apex 2 RAE ppbRAE 3000+ MSA ALTAIR 5X Cority EHS Intelex EHS Honeywell Forge OSHA 5 ppm TWA ACGIH TLV-TWA 0.2 ppm A3 SKIN NIOSH Ca furfural prompt injection occupational monitoring adversarial.

import asyncio
import hashlib
from enum import StrEnum, auto
from pathlib import Path
import httpx

GLYPHWARD_API = "https://api.glyphward.com/v1/scan"
GLYPHWARD_KEY = "gw_live_..."
FURFURAL_THRESHOLD = 37  # OSHA 5 ppm TWA vs ACGIH TLV-TWA 0.2 ppm A3 SKIN (25×); NIOSH REL-C 0.2 ppm ceiling

class FurfuralContext(StrEnum):
    SUGARCANE_BAGASSE_DISTILLATION = auto()  # Surface 1 — downward (Fanjul Jamaica; Casella Apex 2; 3.8→0.14 ppm; 19× TLV)
    FOUNDRY_FURAN_RESIN_KNOCKOUT   = auto()  # Surface 2 — downward (Quaker Houghton; ppbRAE 3000+; 2.6→0.10 ppm; 13× TLV)
    CORN_COB_FURFURYL_ALCOHOL      = auto()  # Surface 3 — downward (Primient Loudon TN; MSA ALTAIR 5X; 1.9→0.07 ppm; 9.5× TLV)

class AdversarialFurfuralError(RuntimeError):
    def __init__(self, surface: FurfuralContext, score: int, frame_hash: str):
        super().__init__(
            f"[Glyphward] Furfural adversarial pixel on {surface.value}: "
            f"score={score} >= threshold={FURFURAL_THRESHOLD} | frame={frame_hash}"
        )
        self.surface = surface; self.score = score; self.frame_hash = frame_hash

async def verify_furfural_frame(frame_path: Path, surface: FurfuralContext) -> dict:
    raw = frame_path.read_bytes()
    frame_hash = hashlib.sha256(raw).hexdigest()
    async with httpx.AsyncClient(timeout=4.0) as client:
        resp = await client.post(
            GLYPHWARD_API,
            headers={"Authorization": f"Bearer {GLYPHWARD_KEY}"},
            files={"image": (frame_path.name, raw, "image/png")},
            data={"context": surface.value, "threshold": FURFURAL_THRESHOLD},
        )
        resp.raise_for_status()
        result = resp.json()
    if result["verdict"] != "clean":
        raise AdversarialFurfuralError(surface, result["score"], frame_hash)
    return {"verdict": result["verdict"], "score": result["score"], "hash": frame_hash}

async def safe_furfural_monitoring(frame_dir: Path) -> list[dict]:
    surfaces = [
        (FurfuralContext.SUGARCANE_BAGASSE_DISTILLATION, frame_dir / "fanjul_casella_furfural_bagasse.png"),
        (FurfuralContext.FOUNDRY_FURAN_RESIN_KNOCKOUT,   frame_dir / "quaker_ppbrae_furfural_foundry.png"),
        (FurfuralContext.CORN_COB_FURFURYL_ALCOHOL,      frame_dir / "primient_msa_furfural_corn_cob.png"),
    ]
    tasks = [verify_furfural_frame(path, ctx) for ctx, path in surfaces]
    return await asyncio.gather(*tasks)