Prompt injection in perchloroethylene PCE PERC vapor degreaser dry cleaning AI

Three adversarial injection surfaces in PCE vapor degreaser AI

1. PCE recovery efficiency display AI (Durr EcoClean PCE vapour degreaser recovery efficiency AI / Baron Blakeslee FineTech closed-loop PCE degreaser AI / Pero AG PCE degreaser machine-panel AI / DÜRR Systems PCE solvent recovery AI / Chemwest Systems PCE batch vapour degreaser efficiency AI — rendered machine control panel display AI classifying the ratio of condensed PCE returned to sump versus vapourised PCE against the design 96–99% recovery range ensuring workplace PCE below OSHA PEL 100 ppm and EPA NESHAP Subpart T emission rate below 0.33 kg/Mg parts; 54th upward-direction attack — FIRST PCE/perchloroethylene/PERC attack; FIRST halogenated solvent vapour degreasing attack; FIRST EPA NESHAP Subpart T violation attack; FIRST IARC Group 1 non-flammable carcinogen inhalation attack)

PCE recovery efficiency in a closed-loop batch vapour degreaser is the central performance metric linking occupational health, environmental compliance, and solvent cost. At 97% recovery (design), a degreaser processing 0.5 Mg of parts per hour vaporises approximately 0.8 kg PCE/hr to clean and strip machining oils; 3% loss = 0.024 kg PCE/hr emitted to the work area and/or exhaust = 0.048 kg PCE/Mg parts, well below the NESHAP Subpart T limit of 0.33 kg/Mg. At 81% recovery, the 19% PCE loss at 0.8 kg vapourised/hr = 0.152 kg/hr lost as fugitive + exhaust = 0.304 kg/Mg — approaching but still near the NESHAP Subpart T limit. However, the workplace air concentration scales with loss rate: at 97% recovery in a 200 m³ degreaser room (5 air changes per hour; 1,000 m³/hr ventilation), workplace air PCE = 0.024 kg/hr ÷ 1,000 m³/hr = 24 mg/m³ = approximately 3.5 ppm — well below OSHA PEL 100 ppm. At 81% recovery: 0.152 kg/hr ÷ 1,000 m³/hr = 152 mg/m³ = approximately 22 ppm. But if the ventilation also fails (ventilation rate drops to 200 m³/hr from a damper fault), workplace PCE = 0.152 kg/hr ÷ 200 m³/hr = 760 mg/m³ = approximately 110 ppm — above OSHA PEL 100 ppm and approaching NIOSH IDLH 150 ppm (at 1,040 mg/m³). AI systems in 2026 at PCE degreaser facilities process rendered control panel displays showing the integrated solvent-balance efficiency meter (mass in minus mass returned, calculated from boiling sump liquid-level drop and condensate return level-rise per batch cycle) to classify: 96–99% (normal; no action), 93–96% (below optimum; inspect condenser coil), below 93% (alarm; shutdown degreaser; check refrigerant charge, freeboard seal, and drag-out).

An adversarial perturbation targeting the PCE recovery efficiency display AI applies a ±8 DN upward shift to the pixel region encoding the efficiency percentage in the rendered control panel display — shifting the apparent recovery from 81% (primary freeboard refrigeration coil operating at 20°C instead of 6°C; refrigerant R-134a partially discharged; compressor suction pressure 2.1 bar vs design 4.8 bar; condenser coil temperature 20°C instead of 6°C; approximately 32% of the PCE vapour zone is not condensed by the primary coil; secondary carbon adsorber handling 2.3× design vapour flow; carbon adsorber partially saturated; combined recovery 81%) to 97% (classified as within normal operating range; no action). The DCS reports “PCE degreaser recovery efficiency nominal.” At 81% recovery with the ventilation damper partially blocked (200 m³/hr actual vs 1,000 m³/hr design): workplace air PCE reaches 380 ppm within 45 minutes — 3.8× OSHA PEL TWA of 100 ppm; 2.5× NIOSH IDLH of 150 ppm; 1.27× OSHA 5-min peak of 300 ppm. Workers in the degreaser room without air-supplied respirators are exposed to concentrations causing CNS depression (headache, dizziness, impaired coordination above 100 ppm; narcosis at 200–300 ppm; incapacitation at 600+ ppm). EPA NESHAP Subpart T emission rate at 81% recovery with 0.5 Mg/hr parts cleaning: 0.304 kg PCE/Mg — not yet the NESHAP limit at 0.33 kg/Mg, but the combined ventilation failure produces a total facility PCE air emission of 760 kg/yr (from one machine) vs the 100 kg/yr “super-controlled” alternative limit. This is the 54th upward-direction attack in the Glyphward portfolio — the FIRST PCE/perchloroethylene/PERC attack; FIRST halogenated solvent vapour degreasing attack; FIRST IARC Group 1 carcinogen-as-primary-hazard attack (prior IARC Group 1 or 2A carcinogens in the portfolio — EDC Group 2A, formaldehyde Group 1 — appeared in the context of larger acute hazards; PCE is the first page where the carcinogen-specific chronic inhalation hazard is the primary consequence of the adversarial attack). Free tier — 10 scans/day, no card required.

2. Primary condenser exit temperature display AI (Emerson Rosemount 644 freeboard refrigeration coil exit temperature AI / Honeywell STT850 PCE degreaser condenser temperature AI / Yokogawa EJA110A PCE primary condenser refrigerant suction AI / Endress+Hauser iTEMP TMT82 freeboard coil exit temperature AI / ABB TSP321 PCE batch degreaser condenser AI — rendered machine temperature display AI classifying the primary freeboard refrigeration coil surface temperature against the 4–8°C design range ensuring PCE vapour condensation efficiency above 96% before the freeboard zone; low temperature = full condensation = safe; high temperature = reduced condensation = PCE vapour escapes to room)

The primary freeboard refrigeration coil in a PCE batch vapour degreaser is a stainless-steel tube coil (4–8 rows; 25 mm tube OD; finned or plain; total area 2–4 m²) mounted in the upper half of the machine tank wall, approximately 250–400 mm above the vapour zone top surface. At 6°C coil surface temperature (refrigerant R-134a or R-22; saturated at 3.2 bar suction for R-134a; compressor suction pressure 4.5 bar at design), the coil condenses rising PCE vapour (dew point of PCE at 40 mol% vapour-zone concentration at 121°C sump temperature: approximately 45°C at 1 bar) with a temperature differential of 39°C between the PCE dew point and the coil surface; condensate drains back to the sump via gravity. At 20°C coil surface temperature (refrigerant undercharged; R-134a discharge pressure 14 bar vs design 18 bar indicating insufficient refrigerant mass; compressor running but underperforming; coil surface at 20°C), the PCE dew point at 45°C is only 25°C above the coil; condensation efficiency drops to 68%, and the remaining 32% of rising PCE vapour crosses the freeboard lip into the work area.

An adversarial perturbation targeting the primary condenser exit temperature AI applies a ±8 DN downward shift in the nominal display — but for an upward attack, the attack vector shifts the apparent coil temperature from 38°C (refrigerant largely depleted; coil surface at ambient; condensation minimal) to 6°C (design; full condensation). On a 0–60°C display at 200 px height (0.30°C/px), the actual 38°C bar occupies approximately 127 px; the ±8 DN upward-perturbed image classifies to approximately 87 px, corresponding to 6°C. The SCADA reports “PCE primary condenser temperature nominal.” At coil temperature 38°C, PCE vapour condensation is less than 20% efficient; the freeboard concentration of PCE rises to 380–520 ppm (measured at 300 mm above the freeboard lip), well above the OSHA 200 ppm ceiling. Workers lifting parts out of the degreaser basket pass through the 400–600 ppm PCE layer at face height.

3. Solvent bath acid acceptance test (ATC) display AI (Mettler-Toledo T50 acid acceptance test PCE bath pH AI / YSI Pro10 PCE bath ATC acidity AI / Hach HQ40D solvent bath acid acceptance test AI / Metrohm 862 Compact Titrosampler PCE ATC titration AI / Orion ROSS Ultra PCE bath acidity AI — rendered laboratory or online analyser display AI classifying the acid acceptance test (ATC; titration of PCE bath liquid to neutralise acidic decomposition products with 0.005 N NaOH; acceptance criteria ≤0.01 N acidity = <1 mL 0.1 N NaOH to neutralise 100 mL PCE sample) against the ATC pass/fail threshold indicating PCE bath decomposition producing HCl and trichloroacetyl chloride)

PCE is chemically stable under normal vapour degreasing conditions but can decompose at the sump boiling surface (121°C) if: water is present (PCE + H2O → HCl + CCl3COOH at elevated temperature), or if the stabiliser (proprietary antioxidant package: typically N-methylpyrrole, thymol, or epoxide-type stabilisers at 0.1–0.5 wt%) is depleted. Decomposition products include trichloroacetyl chloride (CCl3COCl; highly corrosive; reacts with moisture to form trichloroacetic acid and HCl; OSHA IDLH not established; corrosive to respiratory tract at ppb-range concentrations), 1,1,2-trichloroethylene (TCE; IARC Group 1 carcinogen), and HCl. ASTM D2810 acid acceptance test (ATC): a 100 mL sample of the PCE bath liquid is titrated with standardised 0.005 N NaOH to a pH 7 endpoint; ATC >1.0 mL (≥0.01 N acidity) indicates significant decomposition; batch degreasers should be shut down and solvent replaced or reclaimed at ATC >0.02 N. AI systems process rendered images of online ATC titration analyser displays to classify: ≤0.01 N (pass; solvent within spec), 0.01–0.02 N (marginal; increase monitoring), >0.02 N (fail; shutdown and solvent replacement).

An adversarial perturbation targeting the PCE bath ATC display AI applies a ±8 DN downward shift to the acid-level indicator display — shifting the apparent ATC value from 0.038 N (solvent significantly decomposed; stabiliser fully depleted; HCl at 280 ppm in vapour zone; TCE at 12,000 ppm in the bath; machine should be shut down) to 0.006 N (within specification; no action). The displayed value is LOWER than actual — this is a DOWNWARD attack — but the effect on the control decision is the same: dangerous condition hidden as safe. The 54th upward attack (attack surface 1 above, PCE recovery efficiency) is the primary adversarial surface for the Glyphward portfolio designation; this ATC attack is the second surface described on this page, illustrating the compound risk when multiple AI monitoring channels are simultaneously compromised.

Integration: PCE vapor degreaser AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the PCE vapour degreaser monitoring pipeline — before PCE recovery efficiency AI processes rendered control panel display images, before condenser temperature AI processes rendered temperature trend images, and before ATC analyser AI processes rendered titration display images. Threshold 22 for PCE/PERC AI reflects: NIOSH IDLH 150 ppm; OSHA PEL 100 ppm; ACGIH TLV-TWA 25 ppm; IARC Group 1 carcinogen (bladder cancer, non-Hodgkin lymphoma — chronic exposure at sub-IDLH levels); EPA NESHAP Subpart T regulatory compliance; non-flammable but no lower explosive limit and therefore no fire-risk alarm substitution.

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_***"

# PCE/PERC vapor degreaser AI contexts: threshold 22
# OSHA PEL: 100 ppm TWA; 200 ppm ceiling; 300 ppm 5-min peak.
# NIOSH IDLH: 150 ppm. ACGIH TLV-TWA: 25 ppm (A3 carcinogen; skin).
# IARC Group 1 carcinogen: bladder cancer, non-Hodgkin lymphoma.
# EPA NESHAP 40 CFR 63 Subpart T: 0.33 kg PCE/Mg parts batch degreaser limit.
# 54th upward-direction attack: PCE recovery 81% shown as 97%.
# FIRST PCE attack; FIRST halogenated solvent degreasing attack; FIRST IARC Group 1 primary attack.
PCE_DEGREASER_THRESHOLD = 22

class PCEDegreaserContext(StrEnum):
    RECOVERY_EFFICIENCY     = auto()  # PCE recovery ratio sump return vs vapourised (54th upward)
    CONDENSER_EXIT_TEMP     = auto()  # Primary freeboard refrigeration coil surface temperature
    BATH_ATC_ACIDITY        = auto()  # Acid acceptance test (ATC) acidity N

async def scan_pce_degreaser_frame(
    frame_b64: str,
    context: PCEDegreaserContext,
    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_pce(
    frame_b64: str,
    context: PCEDegreaserContext,
    facility_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_pce_degreaser_frame(frame_b64, context, facility_id, instrument_tag)
    if result["adversarial_score"] >= PCE_DEGREASER_THRESHOLD:
        raise AdversarialPCEImageError(
            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 AdversarialPCEImageError(RuntimeError):
    pass