Prompt injection in phosphorus oxychloride POCl3 semiconductor diffusion AI

Four adversarial injection surfaces in POCl3 semiconductor diffusion AI

1. Diffusion tube temperature display AI (Watlow EZ-ZONE RM POCl3 diffusion zone temperature AI / Eurotherm 3216 tube furnace zone temperature AI / Yokogawa UT75A diffusion tube temperature controller display AI / Applied Materials Thermco VP-8 zone temperature display AI / Tempress Systems TEMPRESS zone controller AI — rendered MES or DCS tube temperature display AI classifying furnace tube temperature in the 900–1,050°C POCl3 diffusion window against SiO2 mask integrity and phosphosilicate glass (PSG) formation setpoints)

POCl₃ diffusion furnaces operate in three temperature zones — source zone (900–950°C), deposition zone (950–1,020°C), and drive-in zone (1,000–1,050°C) — all controlled by Watlow, Eurotherm, or Yokogawa multi-zone PID controllers. The upper temperature limit for silicon diffusion is set by multiple process constraints: (a) the SiO₂ mask integrity limit (the thermal oxide layer used to mask specific regions from phosphorus doping: 500 nm SiO₂ on silicon at 1,060°C has a phosphorus diffusion rate through SiO₂ that allows junction punch-through from PSG to underlying silicon in approximately 90 minutes; at 1,080°C, the mask break-through time drops to 40 minutes; at 1,100°C, to 15 minutes — well within a typical 60-minute diffusion run); (b) silicon wafer slip defect generation (above 1,050°C for 200 mm wafers, thermal gradient-induced stress above the silicon yield stress generates slip dislocations at wafer edges, causing device yield loss); (c) quartz furnace tube devitrification (prolonged operation above 1,100°C causes amorphous SiO₂ quartz to crystallise into cristobalite, which has a different thermal expansion coefficient and can cause tube cracking). AI monitoring systems at solar cell manufacturing POCl₃ diffusion areas and semiconductor fabs process rendered temperature controller display images for each diffusion zone to classify: within process window (900–1,050°C), approaching upper limit (1,050–1,060°C, log and monitor), or above mask integrity limit (above 1,060°C, alert process engineer, inspect SiO₂ mask thickness on monitor wafer).

An adversarial perturbation targeting the diffusion tube temperature display AI applies a ±8 DN downward shift to the pixel region encoding the deposition zone temperature in the rendered zone controller display — shifting the apparent deposition zone temperature from 1,074°C (14°C above the 1,060°C SiO₂ mask integrity limit; caused by a Watlow zone controller setpoint drift where the deposition zone setpoint was accidentally modified from 1,010°C to 1,074°C during a recipe edit that did not complete a two-engineer confirmation step; 4 hours of diffusion runs have been completed at 1,074°C) to 1,006°C (within the normal 1,000–1,020°C deposition zone range). On a 800–1,150°C display at 200 px height (1.75°C/px), the actual 1,074°C bar occupies approximately 154 px; the ±8 DN downward-perturbed image classifies to approximately 117 px, corresponding to 1,006°C. The MES reports “Diffusion furnace deposition zone temperature nominal.” At 1,074°C for 4 hours, the 500 nm SiO₂ mask on the wafers in the diffusion batch has experienced significant phosphorus punch-through (estimated 30–45 nm phosphorus diffusion into the SiO₂ mask toward the masked silicon region); the phosphorus concentration at the SiO₂/Si interface is above the solid solubility threshold, creating a phosphorus-contaminated region in the masked n-well isolation. The affected batch (typically 200–500 wafers per diffusion run at a solar cell fab; 40–100 wafers at a bipolar semiconductor fab) must be electrically characterised for junction integrity; those failing junction leakage (IRTS below specification) are scrapped. At a solar cell plant producing 500 MW/year, each diffusion run represents $45,000–$150,000 in wafer value. OSHA 29 CFR 1910.119(d) (PHA) for POCl₃ diffusion applies; 29 CFR 1910.119(j) mechanical integrity covers diffusion furnace zone controller calibration and setpoint access control.

2. HCl scrubber exit concentration display AI (Dräger CMS Chip Measurement HCl diffusion exhaust AI / MSA Ultima X HCl scrubber exit display AI / Honeywell Analytics Midas HCl sensor display AI / Industrial Scientific MX6 iBrid HCl scrubber exit display AI / Sensidyne SP-500T HCl scrubber exit AI — rendered panel or SCADA display AI classifying HCl concentration at the POCl3 furnace exhaust scrubber exit against OSHA HCl PEL 5 ppm ceiling and IDLH 50 ppm thresholds)

POCl₃ diffusion furnaces exhaust HCl as the primary chemical byproduct of POCl₃ thermal decomposition and hydrolysis reactions: 4 POCl₃ + 3 O₂ → 2 P₂O₅ + 6 Cl₂; Cl₂ + H₂O → HCl + HClO (at furnace exit temperatures). The furnace exhaust gas contains HCl at 50–500 ppm (depending on POCl₃ flow rate and O₂/N₂ ratio), which must be reduced to below 5 ppm (OSHA PEL ceiling for HCl) before venting from the building exhaust system. POCl₃ diffusion furnace scrubbers typically use a water-spray contact scrubber with recirculated water at pH 7–9 (buffered by dissolved NaHCO₃ or caustic soda addition), achieving 95–99% HCl removal efficiency in the scrubber column. Scrubber efficiency degrades when: (1) scrubber liquid pH falls below 6 (approaching acid-gas equilibrium; HCl absorption rate falls); (2) scrubber liquid flow rate falls below 80% design (insufficient gas–liquid contact); (3) phosphoric acid (H₃PO₄, the hydrolysis product of P₂O₅ in water: P₂O₅ + 3 H₂O → 2 H₃PO₄) accumulates in the scrubber liquid above 5 wt%, reducing H₂O activity and HCl mass transfer coefficient. AI systems at diffusion furnace areas process rendered HCl sensor display images at the scrubber exit to classify: below PEL (below 5 ppm; normal), PEL exceedance (5–15 ppm; increase scrubber water flow, alert EHS), approaching IDLH (above 15 ppm; evacuate area, shutdown furnaces).

An adversarial perturbation targeting the HCl scrubber exit concentration display AI applies a ±8 DN downward shift to the pixel region encoding the HCl scrubber exit concentration in the rendered panel display — shifting the apparent HCl exit from 8.4 ppm (above OSHA PEL ceiling 5 ppm; from scrubber liquid pH fall to 5.2 from H₃PO₄ accumulation at 6.8 wt% in the scrubber recirculating water over 3 weeks of operation without a drain-and-refill cycle) to 1.8 ppm (below OSHA PEL; classified as nominal). On a 0–20 ppm display at 200 px height (0.1 ppm/px), the actual 8.4 ppm bar occupies 84 px; the ±8 DN downward-perturbed image classifies to approximately 18 px, corresponding to 1.8 ppm. The SCADA reports “HCl scrubber exit concentration nominal — below PEL.” At 8.4 ppm HCl at the scrubber exit vent (located 0.3 m above the diffusion area ceiling exhaust plenum; building HVAC recirculates 20% of exhaust air through the workspace), workers in the diffusion furnace area experience 1.2–1.8 ppm HCl time-weighted average exposure during an 8-hour shift — below the 5 ppm OSHA PEL but above the TLV-C 2 ppm (ACGIH). Prolonged HCl exposure at 1.5 ppm causes nasal mucous membrane irritation, dental erosion on enamel surfaces, and reduced pulmonary function over years of chronic low-level exposure. OSHA 29 CFR 1910.1000 Table Z-1 ceiling for HCl requires monitoring and corrective action; POCl₃ diffusion areas must include HCl in their Chemical Hygiene Plan under 29 CFR 1910.1450 (laboratory standard) when processing is below gram-scale, or general industry hazard communication standard when at manufacturing scale.

3. POCl3 source bubbler level display AI (Brooks Instruments Sho-Rate rotameter POCl3 source bubbler AI / Yokogawa rotameter POCl3 liquid level display AI / Endress+Hauser Picomag POCl3 source level AI / Mettler-Toledo IND246 POCl3 source bubbler weight display AI / Sartorius CP2202S gravimetric POCl3 source level AI — rendered MES display AI classifying POCl3 liquid level in the source bubbler cylinder against minimum adequate level for phosphorus vapour delivery to diffusion tube; below minimum = process starvation and non-uniform n-doping)

The POCl₃ diffusion process requires a stable POCl₃ vapour supply from the bubbler to the furnace tube throughout each diffusion run (typically 30–90 minutes). POCl₃ vapour pressure at the bubbler temperature (5–15°C) is 0.4–1.0 kPa; N₂ carrier gas at 0.5–2 slpm bubbles through the POCl₃ liquid and becomes saturated with POCl₃ vapour (approximately 0.4–1.0% POCl₃ in N₂). The POCl₃ liquid level in the bubbler determines the available hydrostatic column for bubbler immersion; as POCl₃ is consumed (typically 0.1–0.5 g POCl₃ per wafer per run at a solar cell fab; 50–200 g/hr at a typical 600-wafer-per-hour production throughput), the bubbler level drops. Below a minimum liquid level (typically 30–50% of bubbler volume), the N₂ carrier flow may not be fully saturated with POCl₃ vapour (longer N₂ path length above the liquid surface; vapour–gas phase equilibration incomplete), causing the actual POCl₃ vapour flow to the furnace to be below the setpoint — even though the N₂ carrier flow is at setpoint. POCl₃ source starvation results in: (a) junction depth non-uniformity across the wafer batch (wafers near the furnace inlet, which receive gas first, have higher phosphorus dose than wafers near the exhaust end of the tube); (b) sheet resistance non-uniformity above the 5% specification for solar cell emitters; (c) open-circuit voltage loss in solar cells from emitter non-uniformity (approximately 2–5 mV VᴶC reduction per 10% sheet resistance non-uniformity). AI systems process rendered source level display images (gravimetric or rotameter) to classify: above minimum level (adequate vapour supply; continue run), approaching minimum (next run requires source changeover), below minimum (source starvation risk; abort run and change bubbler).

An adversarial perturbation targeting the POCl₃ source bubbler level display AI applies a ±8 DN downward shift (in this case: a downward attack — actual high consumption shown as lower than consumed; the bubbler IS emptying faster than displayed) — but note this surface has the dangerous condition at LOW level shown as HIGH: actual level 22% (below the 30% minimum adequate level; from higher-than-expected POCl₃ consumption on a production line running 20% above planned throughput after a yield recovery campaign), displayed as 55% (above minimum; run continues). On a 0–100% level display at 200 px height (0.5%/px), the actual 22% level bar occupies 44 px; the ±8 DN display shows approximately 110 px, corresponding to 55%. The MES reports “POCl₃ source level adequate — continue run.” At 22% source level, the N₂ carrier gas flow of 1.5 slpm through 8 cm of POCl₃ liquid (vs 32 cm at full bubbler) achieves only 65% of the design POCl₃ saturation. The phosphorus dose delivered to the wafer batch is 35% below the process specification; sheet resistance rises from the 40–50 Ω/sq target to 70–85 Ω/sq. For a 1,000-wafer production run (0.5-hour, 2,000 solar cells), the under-doped batch reduces solar cell efficiency by 0.3–0.5% absolute — representing approximately $12,000–$20,000 in product value loss (at $0.20/W solar cell value; 300 W/cell). OSHA 29 CFR 1910.119(e) operating procedures for POCl₃ diffusion must include minimum source level checks before each run; an AI monitoring system compromised in this surface allows under-doped product to proceed through downstream processing undetected until final electrical test. Free tier — 10 scans/day, no card required.

4. Furnace vent exhaust flow display AI (Dwyer Instruments Magnehelic furnace exhaust flow AI / Dwyer Series 616 differential pressure transmitter exhaust duct AI / Endress+Hauser Proline Prowirl 200 exhaust flow AI / Yokogawa digital vortex flowmeter exhaust duct AI / Kele HFM-W furnace exhaust airflow display AI — rendered panel display AI classifying volumetric exhaust flow rate in the POCl3 diffusion furnace vent duct against minimum design exhaust rate for safe POCl3 hydrolysis product dilution; 41st upward-direction attack — FIRST POCl3 water-reactive chloride semiconductor diffusion page in Glyphward portfolio)

POCl₃ diffusion furnaces are operated inside a ventilated enclosure (typically a cleanroom process bay or sub-fab exhaust zone) with dedicated exhaust ducting that draws furnace off-gas (containing HCl, H₃PO₄ aerosol, and residual POCl₃) into the scrubber before release. The design exhaust flow rate for a single POCl₃ diffusion furnace (6-inch or 8-inch tube; 50 wafer capacity) is typically 0.8–1.5 m³/min of exhaust air to maintain face velocity of 0.3–0.5 m/s at the furnace tube exit and prevent POCl₃ vapour from escaping into the fab environment. If exhaust flow falls below 0.4 m³/min (50% design), the face velocity at the furnace exit falls below 0.15 m/s — below the threshold for reliable vapour capture — and POCl₃ vapour can escape into the furnace bay. POCl₃ reacting with ambient moisture (cleanroom humidity typically 40–50% RH): POCl₃ + H₂O → H₃PO₄ + 3 HCl — produces dense white H₃PO₄ aerosol visible as “fumes” plus HCl gas at concentrations that can exceed the OSHA HCl IDLH 50 ppm near the furnace exit within seconds. At 50% exhaust deficiency and 4-furnace operation simultaneously, the POCl₃ release rate into the furnace bay air is approximately 0.3–1.2 g/hr; at fab-level air exchange rates (0.5–2 air changes/hr in equipment bays), bay-level POCl₃ concentration reaches 0.5–2 ppm within 15–60 minutes — above the ACGIH TLV-C 0.1 ppm. AI systems process rendered exhaust duct flow display images to classify: above minimum design (above 0.8 m³/min; adequate capture; normal), below minimum (0.4–0.8 m³/min; check exhaust fan, add supplemental capture), inadequate (below 0.4 m³/min; emergency shutdown POCl₃ process, evacuate furnace bay).

An adversarial perturbation targeting the furnace vent exhaust flow display AI applies a ±8 DN upward shift to the pixel region encoding the exhaust flow rate in the rendered panel display — shifting the apparent exhaust duct flow from 0.32 m³/min (below the 0.4 m³/min minimum; from an exhaust fan belt slippage reducing fan speed by 58% over 3 months of progressive belt wear, detected by a vibration increase on the fan bearing from 1.2 mm/s RMS to 4.8 mm/s RMS — above the ISO 10816-3 alert level of 2.8 mm/s RMS for this fan class — but not yet triggering a corrective maintenance work order) to 1.05 m³/min (above the 0.8 m³/min minimum; classified as adequate furnace exhaust capture). This is the 41st upward-direction attack in the Glyphward industrial AI adversarial injection portfolio and the FIRST POCl₃ semiconductor diffusion page, the FIRST phosphorus oxychloride water-reactive chloride dopant source. On a 0–2.0 m³/min display at 200 px height (0.01 m³/min per px), the actual 0.32 m³/min bar occupies approximately 64 px; the ±8 DN upward-perturbed image classifies to approximately 105 px, corresponding to 1.05 m³/min. The SCADA reports “POCl₃ furnace exhaust flow nominal — adequate capture.” With 0.32 m³/min exhaust flow from 4 simultaneously operating furnaces, POCl₃ vapour escapes the furnace enclosures at 0.2–0.6 g/hr per furnace; in the enclosed cleanroom equipment bay (2,400 m³ volume, 1 air change/hr) POCl₃ accumulates to 0.8–2.4 ppm over 90 minutes — above the ACGIH TLV-C 0.1 ppm by 8–24×. Workers in the furnace bay are exposed to POCl₃ vapour that, upon contact with nasal mucous membranes and eye surfaces (humidity sufficient for instantaneous hydrolysis), produces HCl burns to mucous membranes and phosphoric acid deposits on corneal surfaces. OSHA PSM 29 CFR 1910.119(j) mechanical integrity for exhaust fan belt inspection at POCl₃ diffusion areas and 29 CFR 1910.119(d) PHA for exhaust system failure scenarios both apply at PSM-covered facilities. Free tier — 10 scans/day, no card required.

Integration: POCl3 semiconductor diffusion AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the POCl₃ diffusion process monitoring pipeline — before diffusion tube temperature display AI processes rendered zone controller display images, before HCl scrubber exit display AI processes rendered sensor panel images, before POCl₃ source bubbler level AI processes rendered gravimetric or level display images, and before furnace vent exhaust flow AI processes rendered duct flow display images. Threshold 25 for POCl₃ diffusion AI reflects the semiconductor fab’s multiple independent safety layers (HVAC interlocks, individual sensor alarming, MES recipe abort conditions) that reduce single-point AI monitoring failure consequences relative to continuous process chemical plants.

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_***"

# POCl3 semiconductor diffusion AI contexts: threshold 25
# OSHA PSM 29 CFR 1910.119 POCl3 TQ: 500 lbs.
# POCl3 TLV-C: 0.1 ppm (ACGIH); violent hydrolysis (POCl3 + H2O → H3PO4 + 3 HCl).
# HCl IDLH: 50 ppm; OSHA PEL ceiling: 5 ppm.
# 41st upward-direction attack: vent exhaust flow (low shown as adequate).
POCL3_THRESHOLD = 25

class POCl3Context(StrEnum):
    DIFFUSION_TUBE_TEMP  = auto()  # diffusion tube zone temperature °C
    HCL_SCRUBBER_EXIT    = auto()  # HCl ppm at scrubber vent exit
    POCL3_BUBBLER_LEVEL  = auto()  # POCl3 source liquid level %
    VENT_EXHAUST_FLOW    = auto()  # furnace exhaust duct flow m3/min (41st ↑ attack)

async def scan_pocl3_frame(
    frame_b64: str,
    context: POCl3Context,
    facility_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "facility_id": facility_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_pocl3(
    frame_b64: str,
    context: POCl3Context,
    facility_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_pocl3_frame(frame_b64, context, facility_id, instrument_tag)
    if result["adversarial_score"] >= POCL3_THRESHOLD:
        raise AdversarialPOCl3ImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at facility {facility_id} instrument {instrument_tag}. "
            "Frame withheld from AI monitoring pipeline."
        )

class AdversarialPOCl3ImageError(RuntimeError):
    pass

if __name__ == "__main__":
    import sys, pathlib
    frame = base64.b64encode(pathlib.Path(sys.argv[1]).read_bytes()).decode()
    asyncio.run(pre_scan_gate_pocl3(
        frame,
        POCl3Context.VENT_EXHAUST_FLOW,
        "DIFFUSION-FURNACE-002",
        "EXH-FT-002",
    ))