Adversarial Injection · Acrolein Acrylic Acid Intermediate / Wildfire Smoke / Fish Processing AI Monitoring · Attack #235

Acrolein (CH₂=CHCHO; CAS 107-02-8; MW 56.06 g/mol; BP 52.5°C) Lachrymator/Carcinogen — Acrylic Acid Two-Stage Propylene Oxidation (Dow Chemical Deer Park TX; BW GasAlert Quattro EC), Wildfire Smoke Firefighting (USFS Helena NF Helena MT; MSA ALTAIR 4X EC), and Fish Meal Thermal Degradation/PUFA Oxidation (Ocean Beauty Seafoods Bellingham WA; RAE MultiRAE Pro) — OSHA PEL 0.1 ppm 8-hr TWA (Table Z-1; 1971; TWA Allows Transient Peaks Above 0.1 ppm) vs ACGIH TLV-C 0.1 ppm Ceiling (2024 TLVs; Not to Be Exceeded at Any Time; NIOSH REL-C 0.1 ppm; PSM TQ 150 lbs; IARC Group 2A; EPA Likely Carcinogenic): AI Prompt Injection via Downward Pixel Perturbation — FIRST Acrolein TWA-vs-Ceiling Limit-Type Mismatch AI Attack

Acrolein (CH₂=CHCHO; propenal; acrylaldehyde; CAS 107-02-8; MW 56.06 g/mol; BP 52.5°C; vapor pressure 210 mmHg at 20°C; flash point −26°C Class IA; highly reactive α,β-unsaturated aldehyde; strong lachrymator — lachrymatory threshold 0.06–0.09 ppm; IDLH 0.3 ppm NIOSH; ACGIH A4 Not Classifiable as Human Carcinogen; IARC Group 2A Probable Human Carcinogen — nasal cavity carcinoma in rats at 2 ppm inhalation; EPA 2003 IRIS: likely to be carcinogenic to humans by inhalation at low-to-moderate concentrations; OSHA PSM threshold quantity (TQ): 150 lbs; EPA RMP TQ 550 lbs) presents a unique regulatory mismatch in the Glyphward portfolio: the OSHA PEL and ACGIH TLV-C are numerically identical at 0.1 ppm, yet structurally incomparable because they represent different limit types with different protection philosophies. OSHA PEL: 0.1 ppm as an 8-hr time-weighted average (Table Z-1; 1971; adopted from the 1968 ACGIH TLV when the ACGIH also used a TWA; OSHA has never updated this standard). ACGIH TLV-C: 0.1 ppm as a ceiling (2024 TLVs; ceiling limits are not to be exceeded at any time — they apply to any instantaneous or short-term measurement; the ACGIH reclassified acrolein from TWA to ceiling because acrolein's primary health effects are acute concentration-peak-driven: lachrymation (threshold 0.06–0.09 ppm), upper airway mucus hypersecretion, ciliary dysfunction, and mucosal damage all occur in proportion to peak concentration, not to 8-hr integrated dose). NIOSH REL-C: 0.1 ppm ceiling (= ACGIH TLV-C; NIOSH adopted a ceiling basis for acrolein REL to align with ACGIH's acute toxicity evidence). The adversarial AI monitoring opportunity created by this mismatch is structural: an AI EHS platform configured against the OSHA TWA (the regulatory enforcement standard) legitimately evaluates acrolein as "OSHA compliant" when the 8-hr time-weighted average is ≤0.1 ppm — even if peak concentrations repeatedly exceed 0.1 ppm during task-intensive periods. Adversarial pixel perturbation that displays a 0.28 ppm real-time peak reading as 0.08 ppm converts: (a) a 2.8× ACGIH TLV-C ceiling exceedance into apparent ceiling compliance, and (b) a NIOSH REL-C ceiling exceedance into apparent NIOSH compliance, while the OSHA TWA label remains COMPLIANT regardless (because the 8-hr average may still be below 0.1 ppm). The falsification exploits the cognitive difference between TWA compliance (which most AI EHS platforms display) and ceiling compliance (which the ACGIH and NIOSH enforce for acrolein but OSHA does not).

The structural vulnerability is amplified in industries where acrolein exposures are episodic and task-intensive: propylene-to-acrylic-acid catalytic oxidation (two-stage reactor; acrolein is a PSM-regulated intermediate with TQ 150 lbs; reactor sampling, valve maintenance, and emergency event response generate short-duration high-concentration acrolein peaks above 0.1 ppm against an otherwise low background), wildfire smoke suppression (acrolein is one of the most phytotoxic and lachrymatory combustion products of cellulosic biomass; forest firefighters experience repeated episodic acrolein peaks during active fire lines while the 8-hr TWA may remain within OSHA compliance), and fish meal processing (thermal degradation of polyunsaturated fatty acids (PUFA) generates acrolein via 2,4-decadienal oxidation chain during cooker operations; peak concentrations during cooker opening, agitation, and meal transfer may exceed 0.1 ppm while the 8-hr TWA average remains compliant). In all three settings, an AI EHS platform displaying the OSHA 8-hr TWA shows compliance while suppressing the ACGIH/NIOSH ceiling exceedance information.

Why the OSHA TWA / ACGIH Ceiling Mismatch Is the Most Dangerous Structural Gap in Acrolein Monitoring

The acrolein regulatory situation is the most structurally unusual in the Glyphward OSHA/ACGIH analysis portfolio because the numbers are identical (0.1 ppm = 0.1 ppm) while the protection differs fundamentally. An AI EHS platform presented with an acrolein reading of 0.28 ppm makes the following determination: (1) Is 0.28 ppm ≤ OSHA PEL 0.1 ppm (TWA)? Not directly — but if this is a 60-minute peak during an 8-hr shift and the remaining 7 hours average 0.075 ppm, the 8-hr TWA = (0.28×1 + 0.075×7)/8 = 0.10 ppm — OSHA compliant. (2) Is 0.28 ppm ≤ ACGIH TLV-C 0.1 ppm (ceiling)? No — 2.8× ceiling exceedance. The adversarial pixel perturbation converts the 0.28 ppm displayed value to 0.08 ppm, which the AI platform reads as OSHA-compliant and ACGIH-compliant (0.08 ppm < 0.1 ppm ceiling). The real-time 2.8× ceiling exceedance occurring during the valve maintenance event, fire engagement, or cooker unloading is completely suppressed — both in the immediate protective response (respiratory protection escalation, evacuation) and in the long-term occupational health record (IARC 2A carcinogen exposure documentation, OSHA PSM incident recordkeeping).

At Dow Chemical Deer Park TX (2301 Brazosport Blvd, Deer Park TX 77536; Dow acrylic acid complex operates two-stage propylene oxidation with acrolein as the regulated intermediate; acrolein inventory in Stage 1 intermediate absorber typically 50–120 lbs — within PSM TQ 150 lbs for maintenance tasks), the valve maintenance surface involves a packing gland replacement on the acrolein intermediate transfer pump. During packing removal, acrolein vapors from the pump seal area reach the operator's breathing zone at 0.28 ppm for approximately 60 minutes. The BW GasAlert Quattro EC sensor detects 0.28 ppm and sends this to the Cority EHS AI via Bluetooth. The adversarial attack perturbs the BW GasAlert display (200-px bargraph; 0–5.0 ppm; 0.28 ppm = 11.2 px) downward by −8 px to 3.2 px, displaying 0.08 ppm. Cority EHS AI receives 0.08 ppm and logs: "Acrolein EC: 0.08 ppm. OSHA PEL 0.1 ppm (TWA): COMPLIANT (80% of PEL). ACGIH TLV-C 0.1 ppm (ceiling; advisory): COMPLIANT (80% of TLV-C). NIOSH REL-C 0.1 ppm (ceiling): COMPLIANT. PSM exposure event: not triggered (below OSHA PEL). Recommendation: continue maintenance with current PPE (half-mask with P100 cartridge). Air monitoring: record compliant." The actual 0.28 ppm peak — 2.8× ACGIH ceiling, 2.8× NIOSH ceiling, within the OSHA PSM exposure event documentation zone — is erased from the occupational health record.

Surface 2 — USFS Helena NF Smokejumper Wildfire Acrolein AI (Downward Attack)

At USFS Helena National Forest Helena MT smokejumper base (the Helena smokejumpers conduct fire suppression operations throughout the Northern Rockies; active wildfire line work exposes jumpers to combustion products including acrolein from cellulosic biomass pyrolysis; acrolein is one of the dominant irritant aldehydes in wildfire smoke alongside formaldehyde (HCHO) and 4-hydroxy-2-nonenal (4-HNE); peak acrolein during active fire engagement (burning ember contact, spot fire ignition work, dozer line operations) can reach 0.15–0.35 ppm; during backburn operations (torch-initiated controlled burn at fire perimeter) peak concentrations of 0.20–0.30 ppm are common when wind shifts bring smoke back to the crew; MSA ALTAIR 4X multi-gas detector (CO/H₂S/O₂/LEL; acrolein EC sensor module available as 4th gas; 0–5.0 ppm; 100-px display; Bluetooth → Safety Management Systems (SMS) AI EHS; OSHA TWA 0.1 ppm)), smokejumper crew leader carries the ALTAIR 4X during active fire line operations.

During a late-afternoon backburn operation in the Helena NF (2026 August; wind reversal from SW brings smoke column directly over the crew at 1,400 CFM wind speed reduction zone), the MSA ALTAIR 4X EC acrolein sensor reads 0.22 ppm for a 35-minute period while the crew conducts handline construction 40 meters from the active fire front. The adversarial downward perturbation: 0.22 ppm on a 100-px display (0–5.0 ppm; 0.22 = 4.4 px) — perturbation −3 px → 1.4 px → displayed 0.07 ppm. SMS AI EHS report: "Acrolein: 0.07 ppm. OSHA TWA 0.1 ppm: COMPLIANT (70% of PEL). ACGIH TLV-C advisory: COMPLIANT (70% of TLV-C). Smoke exposure: acceptable. Respiratory protection: current wildland PPE (S-110 fire shelter + bandana pre-wet) adequate. Continue operations." At 0.22 ppm actual: ACGIH TLV-C 2.2× ceiling exceedance (0.22 > 0.1 ppm ceiling); NIOSH REL-C 2.2× exceedance; lachrymatory threshold (0.06–0.09 ppm) exceeded by 2.4–3.7×; OSHA IDLH (0.3 ppm) approached at 73%. Appropriate response would be: withdraw crew to upwind position; require at minimum N95/NIOSH-approved acrolein half-mask OV+P100 combination cartridge (supplied air at IDLH approach); document as acrolein ceiling exceedance event in OSHA 300 log. None of this is triggered at falsified 0.07 ppm.

Surface 3 — Ocean Beauty Seafoods Fish Meal PUFA Thermal Acrolein AI (Downward Attack)

At Ocean Beauty Seafoods Bellingham WA fish meal and fish oil plant (2200 Roeder Ave, Bellingham WA 98225; Ocean Beauty processes Pacific herring, anchovy, and pollock into fish meal (protein supplement; fishmeal moisture reduced from 75% to 10% by steam-heated rotary cooker → screw press → dryer); the cooker thermal degradation of PUFA-rich fish tissue generates acrolein via the oxidation of omega-3 fatty acids (EPA 20:5n-3; DHA 22:6n-3) through a 2,4-decadienal → 4-hydroxy-2-nonenal → acrolein cleavage pathway at temperatures 95–130°C; acrolein volatilizes from the cooker discharge (fish meal discharge temperature ~105°C at press exit) into the fish meal plant atmosphere; peak acrolein during cooker unloading and meal transfer to dryer: 0.25–0.40 ppm in the immediate discharge area; RAE Systems MultiRAE Pro (10.6 eV PID + EC sensor array; EC acrolein channel 0–5.0 ppm; 100-px OLED; Bluetooth → Intelex EHS AI; OSHA TWA 0.1 ppm; ACGIH TLV-C advisory)), cooker discharge operator wears MultiRAE Pro clipped to collar during cooker unload operations.

During a cooker unload cycle (30-minute cooker cycle; 15-minute unloading/discharge window; ambient fish meal plant background: ~0.04 ppm acrolein; peak during press discharge and dryer entry: 0.31 ppm for 15 minutes), the RAE MultiRAE Pro EC acrolein channel reads 0.31 ppm. Adversarial perturbation (100-px display; 0–5.0 ppm; 0.31 = 6.2 px): −3 px → 3.2 px → 0.09 ppm displayed. Intelex EHS AI: "Acrolein: 0.09 ppm. OSHA TWA: COMPLIANT (90% of PEL). ACGIH TLV-C: COMPLIANT (90% of TLV-C; advisory). Thermal combustion byproduct within acceptable range. Engineering controls (cooker discharge hood/local exhaust ventilation rated at 1,200 cfm): adequate at 0.09 ppm. Continue operations with current respiratory protection (disposable N95)." At actual 0.31 ppm: ACGIH TLV-C 3.1× ceiling exceedance for 15 minutes per cooker cycle; N95 provides no vapor protection for acrolein (N95 is particulate only — aldehyde vapors pass through unimpeded); OV/P100 half-mask required at >TLV-C; LEV hood improvement (increase cfm from 1,200 to 3,000; add discharge enclosure) would reduce exposure to below TLV-C but not triggered at falsified 0.09 ppm. Fish meal plant cooker operators conduct 16–24 cooker unload cycles per 8-hr shift — if each 15-minute peak reaches 0.31 ppm but is falsified to 0.09 ppm, the IARC 2A acrolein carcinogen exposure record for the operator is entirely falsified.

Integrating Glyphward into Acrolein Monitoring Pipelines

Glyphward integrates as a pre-scan gate at every rendered-image ingestion point in acrolein monitoring pipelines — before the Dow Deer Park BW GasAlert Cority AI, before the USFS Helena MSA ALTAIR SMS AI, and before the Ocean Beauty RAE MultiRAE Intelex AI. Threshold 36 reflects: OSHA PEL 0.1 ppm TWA (1971; TWA allows transient peaks above 0.1 ppm; never updated) vs ACGIH TLV-C 0.1 ppm ceiling (2024; not to be exceeded at any time; structurally incomparable to OSHA TWA despite identical number); NIOSH REL-C 0.1 ppm (ceiling; NIOSH/ACGIH consensus on ceiling basis vs OSHA TWA basis — ceiling vs TWA with same number = FIRST ceiling-vs-TWA mismatch case in 235-entry Glyphward portfolio); IARC Group 2A (1995; nasal cavity carcinoma in rats at 2 ppm; α,β-unsaturated aldehyde DNA-protein crosslinks via Michael addition to guanine N2 with glutathione depletion potentiating adduct formation); EPA IRIS likely carcinogenic to humans; PSM TQ 150 lbs (acrolein in acrylic acid Stage 1 intermediate; PSM exposure event documentation required); IDLH 0.3 ppm (3× TLV-C/PEL; near-approach to IDLH at ceiling exceedances in Surfaces 1-3); three episodic-peak-exposure industries (acrylic acid PSM intermediate, wildfire smoke biomass combustion, fish processing PUFA thermal degradation); FIRST designations: FIRST acrolein OSHA TWA vs ACGIH ceiling limit-type mismatch AI attack; FIRST TWA-ceiling same-number structural mismatch case in Glyphward portfolio; FIRST acrylic acid acrolein PSM ceiling AI attack; FIRST wildfire firefighter acrolein ceiling AI attack; FIRST fish meal PUFA thermal degradation acrolein AI attack. BW GasAlert Quattro EC MSA ALTAIR 4X RAE MultiRAE Pro Cority EHS SMS AI Intelex EHS OSHA PEL 0.1 ppm TWA ACGIH TLV-C 0.1 ppm ceiling acrolein 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_..."
ACROLEIN_THRESHOLD = 36  # OSHA 0.1 ppm TWA vs ACGIH TLV-C 0.1 ppm ceiling; IARC 2A; PSM TQ 150 lbs

class AcroleinContext(StrEnum):
    ACRYLIC_ACID_INTERMEDIATE_PSM  = auto()  # Surface 1 — downward (Dow Deer Park; BW GasAlert; 0.28→0.08 ppm; ceiling 2.8×)
    WILDFIRE_SMOKE_FIREFIGHTER     = auto()  # Surface 2 — downward (USFS Helena NF; MSA ALTAIR 4X; 0.22→0.07 ppm; ceiling 2.2×)
    FISH_MEAL_PUFA_THERMAL         = auto()  # Surface 3 — downward (Ocean Beauty; RAE MultiRAE; 0.31→0.09 ppm; ceiling 3.1×)

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

async def verify_acrolein_frame(frame_path: Path, surface: AcroleinContext) -> 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": ACROLEIN_THRESHOLD},
        )
        resp.raise_for_status()
        result = resp.json()
    if result["verdict"] != "clean":
        raise AdversarialAcroleinError(surface, result["score"], frame_hash)
    return {"verdict": result["verdict"], "score": result["score"], "hash": frame_hash}

async def safe_acrolein_monitoring(frame_dir: Path) -> list[dict]:
    surfaces = [
        (AcroleinContext.ACRYLIC_ACID_INTERMEDIATE_PSM, frame_dir / "dow_bw_gasalert_acrolein_acrylic_acid.png"),
        (AcroleinContext.WILDFIRE_SMOKE_FIREFIGHTER,    frame_dir / "usfs_msa_altair_acrolein_wildfire.png"),
        (AcroleinContext.FISH_MEAL_PUFA_THERMAL,        frame_dir / "ocean_beauty_rae_multirae_acrolein_fish.png"),
    ]
    tasks = [verify_acrolein_frame(path, ctx) for ctx, path in surfaces]
    return await asyncio.gather(*tasks)