Adversarial Injection · Industrial Chemical AI Monitoring · Attack #123

n-Butyllithium (n-BuLi) Anionic Polymerization SBR/BR: AI Prompt Injection via ±8 DN Pixel Perturbation — FIRST Organolithium AI Attack

n-Butyllithium (n-BuLi; CAS 109-72-8; GHS Category 1 pyrophoric liquid) ignites spontaneously on contact with moisture or air and is stored as a 1.6 M or 2.5 M solution in hexane — a solvent with flash point −22 °C, LEL 1.1 vol%, and autoignition temperature 225 °C that triggers the OSHA PSM flammable liquid threshold quantity (TQ) of 10,000 lbs under 29 CFR 1910.119 Appendix A. Hexane carries a NIOSH IDLH of 1,100 ppm and an OSHA PEL of 500 ppm (8-hr TWA); n-BuLi itself has no established OSHA PEL, no published AIHA WEEL, and CO₂-based or water-based fire suppression systems are both contraindicated — Class D extinguishers or dry sand are required. A single ±8 DN adversarial pixel perturbation on a rendered DCS display image is sufficient to make a depleted N₂ blanket above an n-BuLi storage vessel appear adequately pressurized, to hide moisture contamination driving hexane-n-BuLi reaction, or to conceal an anionic polymerization exotherm approaching the hexane boiling point. Glyphward detects all three attack surfaces at threshold 44 before any image reaches a downstream AI inference call.

n-BuLi is the industrial initiator of choice for living anionic polymerization of solution styrene-butadiene rubber (SSBR), polybutadiene (BR), and polystyrene (PS) — all conducted in hydrocarbon solvent (hexane or cyclohexane) under rigorously anhydrous, oxygen-free conditions. Albemarle Corporation (Magnolia, AR) operates the world's largest n-BuLi manufacturing capacity; other producers include Chemetall GmbH (Frankfurt, Germany, now an Orano subsidiary) and FMC Lithium. Global SSBR demand exceeds 1.5 million tonnes per year driven by high-performance tire tread compounds. n-BuLi reacts violently with water (C₄H₉Li + H₂O → C₄H₁₀ + LiOH; butane evolution plus heat; vigorous H₂ generation under acidic conditions) and with CO₂ (C₄H₉Li + CO₂ → C₄H₉CO₂Li; lithium pentanoate formation, rendering CO₂ suppression systems not merely ineffective but reactive). The UCLA laboratory fatality of Sheri Sangji in January 2009 — caused by tert-butyllithium ignition during a transfer procedure — defined the modern standard for organolithium safe handling in OSHA enforcement guidance and subsequent Cal/OSHA prosecution; every subsequent organolithium OSHA audit, Process Hazard Analysis, and safe work procedure cites that incident. The PSM pyrophoric liquid threshold of 10,000 lbs means a single n-BuLi storage tank holding approximately 1,815 US gallons of 2.5 M hexane solution crosses the PSM threshold.

Why n-BuLi Anionic Polymerization Is Disproportionately Vulnerable to Pixel Manipulation

Three structural features of n-BuLi anionic polymerization make it exceptionally susceptible to adversarial DCS image attacks. First, the process operates under extreme moisture and oxygen exclusion conditions — all n-BuLi handling and polymerization steps require water content below 10–20 ppm in every solvent, monomer, and equipment surface. This means the instruments most critical to safety (N₂ blanket pressure transmitters, Karl Fischer moisture analyzers, and reactor temperature sensors) display values in narrow absolute ranges where small pixel shifts represent large fractional deviations from safety-critical setpoints. A 200-pixel DCS bar spanning 0 to 0.5 bar gauge for the N₂ blanket means each pixel represents only 0.0025 bar; an adversarial shift of 76 pixels — visually subtle on a full dashboard — represents 0.19 bar, fully inverting the safety picture from "blanket failing" to "blanket adequate." AI-based process monitoring systems that classify rendered instrument displays as "normal" or "alarm" are vulnerable to exactly this scale of manipulation.

Second, the consequence of moisture contamination is not linear but threshold-driven. At moisture levels below ~5 ppm, n-BuLi in hexane is stable over days-long storage. Above ~20–30 ppm moisture, initiator deactivation and LiOH precipitation begin; above ~50 ppm, butane evolution becomes a measurable source of overpressure and exotherm. The Karl Fischer moisture display therefore has a narrow "warning" band that the adversarial attack exploits: shifting the displayed reading from 68 ppm (above the action threshold) to 10 ppm (below all alarm thresholds) completely suppresses the response chain (molecular sieve column regeneration, feed tank draining, and instrument notification) that would otherwise prevent n-BuLi degradation and potential fire.

Third, anionic polymerization exotherms are not self-limiting — unlike radical polymerization, there is no natural termination mechanism in living anionic systems, and the propagation rate increases with temperature in a positive feedback loop. The n-BuLi addition phase is the highest-exotherm period of SSBR polymerization: as initiator is added, every chain-end lithium carbanion attacks monomer immediately and irreversibly, releasing approximately 65–85 kJ/mol per styrene or butadiene addition. Heat removal depends entirely on the reactor jacket and reflux condenser. A reactor temperature display showing 14 °C when the actual temperature is 52 °C — only 17 °C below the hexane boiling point of 69 °C — means the AI monitoring system sees no urgency for emergency cooling while the actual process is one further exotherm step away from hexane vapor generation, pressure relief device actuation, and vapor cloud formation in a workspace where ignition sources (pumps, instrumentation, lighting) are present despite best-practice electrical area classification.

Surface 1 — N₂ Blanket Overpressure Above n-BuLi Storage Vessel (Downward Attack)

The N₂ blanket overpressure indicator for the n-BuLi storage vessel spans 0 to 0.5 bar gauge on a 200 px vertical bar in the DCS overview tile. The pixel scale is 200 px ÷ 0.5 bar = 400 px/bar. At the actual blanket pressure of 0.12 bar gauge, the rendered pixel position is 0.12 × 400 = 48 px from the bottom of the bar. The adversarial perturbation shifts this pixel cluster upward by 76 px to position 124 px. The AI inference engine reads the pressure as 124 ÷ 400 = 0.31 bar gauge — 2.6× the actual value. Actual (0.12 bar) is below displayed (0.31 bar): this is a downward attack, inflating the apparent reading to conceal that the blanket is failing.

An n-BuLi storage vessel (typically a jacketed, nitrogen-blanketed stainless steel tank, internally dried to <5 ppm H₂O, maintained at 0.05–0.30 bar gauge N₂ overpressure to prevent air/moisture ingress through seals and sample valves) depends on continuous N₂ supply to maintain positive pressure. When the blanket pressure drops toward atmospheric or below, the pressure differential reverses momentarily during level drawdown or temperature cycling, allowing humid ambient air to enter the tank vapor space. At 0.12 bar — the actual pressure that the attack conceals — the blanket is below the design setpoint minimum of approximately 0.15–0.20 bar; the N₂ top-up valve (controlled by the pressure transmitter via a PLC loop) should have already activated. The displayed 0.31 bar convinces the AI monitoring system that the blanket is operating in the normal mid-range; no N₂ top-up alarm is generated; no operator notification is sent; and the N₂ supply system is not inspected. Moisture and oxygen ingress from even brief atmospheric contact with n-BuLi solution causes immediate surface reaction: the n-BuLi at the gas-liquid interface oxidizes (C₄H₉Li + O₂ → C₄H₉OOLi, lithium butylperoxide; a shock-sensitive compound) and hydrolyzes (C₄H₉Li + H₂O → C₄H₁₀ + LiOH), generating butane gas bubbles, LiOH crust on the liquid surface, and localized exotherms. In worst-case scenarios where the tank seal fails and bulk air contacts the n-BuLi solution surface, spontaneous pyrophoric ignition occurs — a self-sustaining fire that cannot be suppressed with CO₂ or water.

Surface 2 — Hexane Moisture Content in n-BuLi Solution Feed Tank (Upward Attack)

The Karl Fischer moisture content indicator for the hexane feed to the n-BuLi solution tank spans 0 to 100 ppm H₂O on a 200 px vertical bar. The pixel scale is 200 px ÷ 100 ppm = 2 px/ppm. At the actual moisture content of 68 ppm, the rendered pixel position is 68 × 2 = 136 px from the bottom of the bar. The adversarial perturbation shifts this pixel cluster downward by 116 px to position 20 px. The AI inference engine reads the moisture as 20 ÷ 2 = 10 ppm — a 6.8× underread. Actual (68 ppm) is above displayed (10 ppm): this is an upward attack, suppressing the apparent reading to conceal that moisture contamination has reached a dangerous level.

The hexane solvent used in n-BuLi anionic polymerization is specification-grade "anhydrous hexane" (<20 ppm H₂O; dried over activated 4Å molecular sieves; monitored continuously by Karl Fischer coulometric titration at the feed tank outlet). The molecular sieve beds require periodic regeneration (thermal desorption at 250–350 °C under N₂ purge) to restore drying capacity; the regeneration interval is determined by the online moisture analyzer reading. When the display shows 10 ppm — below the 20 ppm regeneration trigger — the AI system concludes the molecular sieve is performing adequately and no regeneration is needed. At the actual moisture level of 68 ppm, the molecular sieve bed has been saturated and is no longer drying the hexane solvent stream effectively: moisture at 68 ppm is entering the n-BuLi solution feed tank continuously. The consequence sequence: n-BuLi reacts with dissolved moisture (C₄H₉Li + H₂O → C₄H₁₀ + LiOH) throughout the solution volume, progressively deactivating the initiator (reducing effective n-BuLi concentration and shifting the polymer molecular weight distribution); LiOH precipitates as fine solid particles that can plug filter elements, sample valves, and reactor transfer lines; butane (bp −1 °C) is evolved as dissolved gas, building pressure in the feed tank vapor space; and the deactivation reaction is exothermic, raising the solution temperature. In a large feed tank (5,000–20,000 liter capacity), even 1–2 °C of exotherm from n-BuLi deactivation represents significant energy release that can propagate if ignition sources are present.

Surface 3 — Polymerization Reactor Temperature During n-BuLi Addition (Upward Attack)

The polymerization reactor temperature indicator during the n-BuLi addition phase spans −10 °C to 60 °C on a 200 px vertical bar. The pixel scale is 200 px ÷ 70 °C = 2.857 px/°C. At the actual reactor temperature of 52 °C, the rendered pixel position is (52 − (−10)) × 2.857 = 62 × 2.857 = 177.1 px from the bottom of the bar. The adversarial perturbation shifts this pixel cluster downward by 109 px to position 68.1 px. The AI inference engine reads the temperature as 68.1 ÷ 2.857 + (−10) = 23.8 − 10 = 13.8 °C, displayed as approximately 14 °C — a 38 °C underread. Actual (52 °C) is well above displayed (14 °C): this is an upward attack, suppressing the reading to conceal that the reactor is running dangerously hot.

During n-BuLi addition to the styrene/butadiene monomer solution (typically a 15–25 wt% monomer in hexane; reactor charge 10,000–50,000 liters; n-BuLi added as a concentrated hexane solution via a metered feed pump), every mole of active chain-end lithium carbanion reacts immediately with monomer in the propagation step, releasing 65–85 kJ/mol of polymerization heat. Reactor temperature is controlled by jacket cooling water (chilled water at 5–15 °C) and by a reflux condenser at the reactor overhead that condenses hexane vapor and returns it to the reactor (hexane bp 69 °C; reflux condensation is the primary heat removal mechanism in many SSBR reactor designs). At an actual reactor temperature of 52 °C, the margin to hexane boiling is only 17 °C; the reflux condenser is approaching its maximum heat removal rate; and if the exotherm rate from the n-BuLi initiation step exceeds the combined cooling capacity of jacket plus condenser, the temperature will rise to 69 °C in minutes. The AI monitoring system, reading 14 °C from the manipulated display, interprets the reactor as cold-side-biased — actually at the low end of the normal operating range — and does not trigger the emergency actions that the true temperature of 52 °C would mandate: emergency cooling water increase, n-BuLi addition pump shutdown, and SIS (Safety Instrumented System) interlock evaluation. Once hexane reaches its boiling point (69 °C), vapor generation accelerates, the rupture disk on the reactor vent header opens (typically set at 1.5–2.0 bar gauge above normal operating pressure), and hexane vapor discharges through the vent to the atmosphere or to a condenser. Hexane flash point is −22 °C: any ignition source — electrical arcing, hot equipment surfaces, static discharge from the vapor stream — can ignite the hexane vapor cloud, resulting in a flash fire or vapor cloud explosion in the polymerization bay. Hexane CERCLA RQ is 5,000 lbs; a 50,000-liter SSBR reactor charged with 15 wt% hexane (approximately 33,000 kg hexane) represents 73,000 lbs of CERCLA-reportable hexane in a single process vessel.

Integrating Glyphward into n-BuLi Anionic Polymerization AI Monitoring Pipelines

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the n-BuLi anionic polymerization AI monitoring pipeline — before the N₂ blanket pressure AI processes rendered pressure transmitter DCS display images, before the hexane moisture AI processes rendered Karl Fischer analyzer display images, and before the reactor temperature AI processes rendered thermocouple DCS display images. All three scan calls execute in parallel via asyncio.gather, keeping total pre-scan latency under 200 ms for a three-channel frame bundle — well within the 1-second inference latency budget of typical process AI monitoring systems.

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_***"

# n-BuLi anionic polymerization AI contexts — threshold 44
# GHS Category 1 pyrophoric liquid; hexane PSM TQ 10,000 lbs; flash point −22 °C.
# n-BuLi reacts with H₂O and CO₂: water and CO₂ suppression both contraindicated.
# UCLA Sangji 2009 organolithium anchor; OSHA enforcement guidance reference.
# Attack #123: FIRST n-BuLi AI attack; FIRST anionic polymerization AI attack;
# FIRST organolithium N₂ blanket pixel-attack; FIRST hexane moisture AI concealment attack.
NBULI_GLYPHWARD_THRESHOLD = 44

class OrganolithiumContext(StrEnum):
    N2_BLANKET_PRESSURE       = auto()  # N₂ overpressure above n-BuLi storage vessel (downward attack)
    HEXANE_MOISTURE_CONTENT   = auto()  # Karl Fischer H₂O in hexane feed tank (upward attack)
    POLYMERIZATION_REACTOR_TEMP = auto()  # reactor °C during n-BuLi addition phase (upward attack)

async def scan_nbuli_frame(
    frame_b64: str,
    context: OrganolithiumContext,
    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_nbuli(
    n2_frame_b64: str,
    moisture_frame_b64: str,
    temp_frame_b64: str,
    plant_id: str,
) -> None:
    results = await asyncio.gather(
        scan_nbuli_frame(n2_frame_b64,       OrganolithiumContext.N2_BLANKET_PRESSURE,        plant_id, "PT-N2-BLANKET-001"),
        scan_nbuli_frame(moisture_frame_b64,  OrganolithiumContext.HEXANE_MOISTURE_CONTENT,    plant_id, "AT-KF-MOISTURE-001"),
        scan_nbuli_frame(temp_frame_b64,      OrganolithiumContext.POLYMERIZATION_REACTOR_TEMP, plant_id, "TT-REACTOR-001"),
    )
    for result, context in zip(
        results,
        [OrganolithiumContext.N2_BLANKET_PRESSURE,
         OrganolithiumContext.HEXANE_MOISTURE_CONTENT,
         OrganolithiumContext.POLYMERIZATION_REACTOR_TEMP],
    ):
        if result["adversarial_score"] >= NBULI_GLYPHWARD_THRESHOLD:
            raise AdversarialNBuLiImageError(
                f"Adversarial injection detected in {context} "
                f"(score {result['adversarial_score']} >= threshold {NBULI_GLYPHWARD_THRESHOLD}) "
                f"at plant {plant_id}. "
                "Frame withheld from n-BuLi anionic polymerization AI pipeline."
            )

class AdversarialNBuLiImageError(RuntimeError):
    pass

Each Glyphward scan call returns an adversarial score, a perturbation hash, and a signed audit record within the 4-second HTTP timeout. The signed audit record — keyed to the SHA-256 hash of the raw frame bytes — provides a cryptographically anchored chain of custody for each DCS image processed by the n-BuLi monitoring AI, satisfying the OSHA PSM 29 CFR 1910.119(j) mechanical integrity recordkeeping requirement for safety-instrumented function documentation. A score at or above threshold 44 blocks the frame from reaching the downstream anionic polymerization AI model and triggers an operator notification via the plant DCS alarm system.