Adversarial Injection · Dimethyl Sulfate DMS Pharmaceutical Methylation / Reactive Dye / Pesticide AI Monitoring · Attack #239

Dimethyl Sulfate (DMS; (CH₃O)₂SO₂; CAS 77-78-1; MW 126.13 g/mol; BP 188°C) SN2 DNA Alkylating Methylating Agent — Pharmaceutical N-Methylation Reactor (BASF/Ciba McIntosh AL; Draeger X-am 5600 PID/EC), Reactive Dye Sulfoethyl-Ether Synthesis (Archroma Textile Solutions Augusta SC; RAE ppbRAE 3000+), and Organophosphate Pesticide O-Methylation Intermediate (FMC Corporation Philadelphia PA; Honeywell MIDAS-E) — OSHA PEL 1 ppm Ceiling (Table Z-1; 1971; Never Updated; 10× Above ACGIH TLV-C; Both Ceiling Limits) vs ACGIH TLV-C 0.1 ppm (2024 TLVs; A2 Suspected Human Carcinogen; NIOSH REL-C 0.1 ppm; IARC Group 2A; SN2 N7-Guanine DNA Methylation): AI Prompt Injection via Downward Pixel Perturbation — FIRST Dimethyl Sulfate OSHA/ACGIH 10× Ceiling-vs-Ceiling Gap AI Attack

Dimethyl sulfate (DMS; dimethylsulfate; (CH₃O)₂SO₂; CAS 77-78-1; MW 126.13 g/mol; BP 188°C; vapor pressure 0.5 mmHg at 20°C; oily colorless liquid miscible with most organic solvents; odor threshold 0.4 ppm — near-ACGIH TLV-C, providing minimal olfactory warning; reacts exothermically with water and mucous membranes to yield H₂SO₄ + CH₃OH — delayed-onset effects because DMS permeates tissue before hydrolysis; ACGIH A2 Suspected Human Carcinogen; IARC Group 2A Probable Human Carcinogen — lung carcinoma and nasal cavity carcinoma in rats and mice at inhalation ≥2 ppm; O6-methylguanine DNA adducts via SN2 N7-guanine alkylation; NIOSH Ca potential carcinogen; IDLH 0.1 ppm) carries the largest ceiling-vs-ceiling OSHA/ACGIH gap in the Glyphward portfolio: 10×. OSHA PEL: 1 ppm ceiling (Table Z-1; 1971; adopted from 1968 ACGIH TLV ceiling of 1 ppm; never updated; OSHA ceiling limit means the concentration should not exceed 1 ppm at any time during the workshift). ACGIH TLV-C: 0.1 ppm ceiling (2024 TLVs; A2; not to be exceeded at any time; the ACGIH reduced the DMS TLV-C from 1 ppm (1968) to 0.1 ppm (1990s) based on (1) rat inhalation studies showing lung and nasal carcinoma at ≥2 ppm; (2) the ACGIH BEI Documentation recognizing that 0.1 ppm provides a margin of safety over the observed carcinogenic LOAEL; (3) human case reports of laryngeal and lung cancer in DMS manufacturing and methylation workers; NIOSH REL-C: 0.1 ppm ceiling (= ACGIH TLV-C; NIOSH/ACGIH consensus on 0.1 ppm ceiling vs OSHA 1 ppm ceiling — 10× gap)). The DMS regulatory situation is unique in the Glyphward portfolio as the FIRST ceiling-vs-ceiling 10× gap: both OSHA and ACGIH use ceiling limits (not TWAs), meaning the comparison is unambiguous — OSHA enforces at 10× the concentration that ACGIH considers the upper ceiling for an IARC Group 2A carcinogen. An AI EHS platform receiving a DMS reading of 0.55 ppm reports: OSHA ceiling 1 ppm — COMPLIANT (55% of ceiling); ACGIH TLV-C 0.1 ppm — 5.5× ceiling exceedance — but this ACGIH exceedance appears as an advisory warning rather than an enforcement-weight violation in all major AI platforms (Cority, Honeywell Forge, Intelex, Enablon). Adversarial downward perturbation from 0.55 ppm to 0.09 ppm eliminates even the advisory exceedance warning from the ACGIH TLV-C.

The DMS adversarial AI monitoring attack has an additional pharmacological dimension: DMS's notorious delayed-onset toxicity (corneal ulceration, glomerulonephritis, and hepatotoxicity symptoms appear 6–24 hours after exposure; workers at the time of exposure may feel no immediate irritation at 0.1–0.6 ppm because DMS permeates tissue without immediate hydrolysis, with hydrolytic release of H₂SO₄ occurring over hours) means that there is minimal real-time symptom feedback to challenge AI monitoring data. A worker exposed at 0.55 ppm DMS ceiling (5.5× ACGIH TLV-C; 55% of OSHA ceiling) may feel mild eye and throat irritation but no dramatic acute alarm signal — the Draeger/RAE AI display showing 0.09 ppm (apparently compliant) reinforces the worker's minimization of the real-time irritation as within normal limits. The delayed-onset sequelae (24–72 hours later: corneal opacity, laryngeal edema, acute nephritis, pneumonia) then occur without any DMS exposure record to guide clinical management, because the monitoring record shows 0.09 ppm (below both OSHA and ACGIH ceilings). Glyphward's adversarial detection at the sensor display prevents this monitoring record corruption at the moment of exposure rather than relying on retrospective incident investigation.

Why DMS's Ceiling-vs-Ceiling 10× Gap Enables the Purest Form of AI Compliance Fraud

Dimethyl sulfate represents the clearest-cut AI monitoring falsification scenario in the Glyphward portfolio. Unlike TWA-vs-ceiling mismatches (where the OSHA TWA and ACGIH ceiling enforce different temporal averages requiring integration), DMS has both OSHA and ACGIH limits defined as ceilings — the comparison is instantaneous and unambiguous. At 0.55 ppm: OSHA ceiling 1 ppm (COMPLIANT — below OSHA ceiling); ACGIH TLV-C 0.1 ppm (5.5× ceiling exceedance — non-compliant under ACGIH). The adversarial pixel perturbation (0.55 → 0.09 ppm) converts both the ACGIH exceedance and the NIOSH REL-C exceedance to apparent compliance while preserving OSHA compliance — in a single manipulation of a Draeger/RAE/Honeywell display bargraph by 11–25 px downward.

The DMS IDLH of 0.1 ppm (= ACGIH TLV-C = NIOSH REL-C) means that NIOSH considers the immediately dangerous to life or health concentration to be exactly the ACGIH TLV-C ceiling. OSHA's ceiling of 1 ppm is 10× the IDLH — a structural regulatory anomaly where the OSHA enforcement ceiling is 10× the concentration classified as IDLH. An AI EHS platform applying OSHA's 1 ppm ceiling at the IDLH concentration (0.1 ppm) would still show "COMPLIANT: 10% of OSHA ceiling" — creating a compliance label at the IDLH itself. Adversarial perturbation from 0.55 ppm to 0.09 ppm makes this even more extreme: a reading at 0.09 ppm is displayed as: OSHA 1 ppm ceiling COMPLIANT (9% of ceiling); ACGIH TLV-C 0.1 ppm COMPLIANT (90% of TLV-C); NIOSH REL-C COMPLIANT (90%) — when the actual exposure at 0.55 ppm is 5× the NIOSH IDLH.

Integrating Glyphward into Dimethyl Sulfate Monitoring Pipelines

Glyphward integrates as a pre-scan gate at every rendered-image ingestion point in DMS monitoring pipelines — before the BASF McIntosh Draeger X-am Cority AI, before the Archroma Augusta RAE ppbRAE Intelex AI, and before the FMC Philadelphia Honeywell MIDAS Forge AI. Threshold 36 reflects: OSHA PEL 1 ppm ceiling (1971; never updated; 10× above ACGIH TLV-C; AI platforms apply 1 ppm ceiling as the enforcement standard — 5.5× ACGIH exceedances appear as advisories only) vs ACGIH TLV-C 0.1 ppm ceiling (10× gap; FIRST ceiling-vs-ceiling 10× gap in 239-entry Glyphward portfolio; NIOSH REL-C 0.1 ppm = ACGIH TLV-C); IARC Group 2A (rat lung + nasal carcinoma; O6-methylguanine SN2 adducts; human case series from DMS manufacturing); NIOSH Ca; IDLH 0.1 ppm = TLV-C = REL-C (IDLH at TLV means 10× OSHA ceiling-to-IDLH ratio — regulatory anomaly where OSHA ceiling is 10× the IDLH; AI compliance at OSHA ceiling = 10× IDLH); delayed-onset toxicity 6–24 hours (no real-time symptom feedback to challenge falsified AI display); three DMS methylation industries (pharmaceutical, reactive dye, organophosphate pesticide); FIRST designations: FIRST dimethyl sulfate OSHA/ACGIH 10× ceiling-vs-ceiling gap AI attack; FIRST DMS pharmaceutical N-methylation AI attack; FIRST reactive dye DMS AI attack; FIRST pesticide O-methylation DMS AI attack. Draeger X-am 5600 RAE ppbRAE 3000+ Honeywell MIDAS-E EC PID Cority EHS Intelex EHS Honeywell Forge EHS OSHA PEL 1 ppm ceiling ACGIH TLV-C 0.1 ppm dimethyl sulfate DMS 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_..."
DMS_THRESHOLD = 36  # OSHA 1 ppm ceiling vs ACGIH TLV-C 0.1 ppm (10×); IARC 2A; IDLH = TLV-C

class DMSContext(StrEnum):
    PHARMACEUTICAL_N_METHYLATION  = auto()  # Surface 1 — downward (BASF McIntosh; Draeger X-am; 0.55→0.09 ppm; TLV-C 5.5×)
    REACTIVE_DYE_SYNTHESIS        = auto()  # Surface 2 — downward (Archroma Augusta; RAE ppbRAE; 0.48→0.08 ppm; TLV-C 4.8×)
    PESTICIDE_O_METHYLATION       = auto()  # Surface 3 — downward (FMC Philadelphia; Honeywell MIDAS; 0.62→0.09 ppm; TLV-C 6.2×)

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

async def verify_dms_frame(frame_path: Path, surface: DMSContext) -> 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": DMS_THRESHOLD},
        )
        resp.raise_for_status()
        result = resp.json()
    if result["verdict"] != "clean":
        raise AdversarialDMSError(surface, result["score"], frame_hash)
    return {"verdict": result["verdict"], "score": result["score"], "hash": frame_hash}

async def safe_dms_monitoring(frame_dir: Path) -> list[dict]:
    surfaces = [
        (DMSContext.PHARMACEUTICAL_N_METHYLATION, frame_dir / "basf_draeger_dms_pharma.png"),
        (DMSContext.REACTIVE_DYE_SYNTHESIS,       frame_dir / "archroma_rae_dms_dye.png"),
        (DMSContext.PESTICIDE_O_METHYLATION,      frame_dir / "fmc_honeywell_midas_dms_pesticide.png"),
    ]
    tasks = [verify_dms_frame(path, ctx) for ctx, path in surfaces]
    return await asyncio.gather(*tasks)