Prompt injection in jet engine borescope inspection AI

Four adversarial injection surfaces in jet engine borescope AI

1. High-pressure turbine (HPT) blade crack and oxidation detection AI

The high-pressure turbine (HPT) stage — operating at gas temperatures of 1,600–1,800°C and blade rotational speeds of 10,000–20,000 RPM in current high-bypass turbofan engines — is the highest-stress component in the engine and the primary target of borescope inspection AI. HPT blades are single-crystal nickel superalloy castings with internal cooling channels and thermal barrier coatings (TBC — typically yttria-stabilised zirconia, YSZ, applied by air plasma spray or electron-beam physical vapour deposition) that limit blade metal temperature to within the nickel superalloy’s creep resistance limit. Borescope inspection of the HPT stage produces rendered video frames and still images of each blade airfoil in the aft-looking-forward (ALF) and forward-looking-aft (FLA) orientations, capturing the leading edge, trailing edge, suction surface, and pressure surface of each blade as it is rotated through the borescope port viewing angle. The HPT AI classification model — trained on annotated datasets from OEM test cell teardowns and field removal inspections — classifies each blade image for: transverse cracks (radial cracks perpendicular to the blade span, high-severity), longitudinal cracks (parallel to the chord, moderate-severity), TBC spallation (loss of the thermal barrier coating exposing the metal substrate, high-severity if >15mm²), leading edge erosion (rounded leading edge from particulate erosion in desert or volcanic ash environments), and platform cracking (cracks in the blade root platform area adjacent to the cooling air flow path). The OEM Engine Maintenance Manual (EMM) defines the serviceable limits for each defect type — the dimensional thresholds beyond which continued service requires OEM engineering authority disposition or mandatory shop removal; the AI classification drives automatic comparison against these EMM limits and generates a disposition flag (serviceable, monitor-at-next-inspection, or remove-now).

An adversarial perturbation on a borescope blade image that suppresses the transverse crack signature — smoothing the apparent dark linear discontinuity across the blade airfoil — can cause the HPT AI to classify a crack-present blade as serviceable, preventing the remove-now flag that would take the engine off wing for shop inspection. The NTSB accident investigation history includes several cases where undetected HPT crack propagation led to blade separation: American Airlines Flight 96 (1972, DC-10 cargo door and engine failure from blade separation), United Airlines Flight 232 (1989, DC-10 fan disk explosion killing 111 of 296 passengers), and Qantas Flight QF32 (2010, A380 Rolls-Royce Trent 970 HPT blade disc failure). The QF32 event — where a Trent 970 HPT disc released an uncontained blade fragment that penetrated the wing fuel tank, hydraulic systems, and flight control systems — was preceded by manufacturing defects (non-conforming oil pipe bore) that the scheduled borescope inspection program did not detect; an adversarially suppressed AI classification in the same inspection pipeline would produce an identical outcome from a software layer rather than a manufacturing layer.

2. Foreign object debris (FOD) ingestion damage classification AI

FOD ingestion events — bird strikes, ice crystal ingestion at cruise altitude, runway debris ingestion on takeoff — trigger mandatory borescope inspection of the fan, LPC (low-pressure compressor), HPC, and HPT stages under FAA Airworthiness Directive requirements and EASA airworthiness directives issued under 14 CFR Part 39 and Part 21 Subpart M. The FOD borescope AI classifies the extent and severity of FOD impact damage on compressor and turbine blade airfoils: dents (surface depressions from low-velocity debris impact), nicks (sharp-edged surface defects from stone or ice impact), tears (material separation at blade leading or trailing edge), and blade erosion profile deviation from nominal chord profile. The FAA Technical Standard Order TSO-C199 for borescope inspection equipment and FAA Order 8900.1 Volume 6 Chapter 9 Section 12 (Air Carrier Operations: Maintenance and Airworthiness) establish the FAA Surveillance Officer review requirements for borescope inspections; EASA AMC 20-16 establishes the approved maintenance AI tool criteria under which AI-assisted borescope inspection classifications must be validated against human inspector review samples. The FOD borescope AI output feeds the “serviceable limits comparison” that drives the inspection disposition in the MRO’s maintenance management system (MRO software: AMOS, TRAX, Swiss AviationSoftware, OASES), which generates the Aircraft Maintenance Release (AMR) that allows the aircraft to return to service under 14 CFR 43.9 maintenance record requirements.

An adversarial perturbation on a FOD-damaged fan blade borescope image that rounds the apparent sharp edge of a nick — shifting the pixel gradient at the blade leading edge from the angular “nick” class toward the smooth “nominal profile” class — can cause the FOD AI to classify a nick-present blade as undamaged, preventing the AMR hold that would require dimensional measurement and engineering disposition before return to service. The A320neo CFM LEAP-1A fan blade icing event (Air India Express Flight IX-611 type, 2019 AAIB bulletin ALS-2019-15) demonstrated how ice crystal ingestion damage to fan blades, if not correctly classified as non-serviceable at borescope, permits return to service of an engine with damage that can propagate to fan blade separation under subsequent thermal cycle loading — an uncontained engine failure event at the energy level of a 5-kJ projectile per separated fan blade (GE Aviation fan blade out test data).

3. Combustor liner hot spot and burn-through detection AI

Combustor liner inspection — the annular or can combustion chamber liner that must contain the 1,600°C combustion gas stream without hot-streak damage to the downstream HPT — uses borescope AI to classify liner panel condition for: hot spots (discolored regions indicating local TBC failure and metal substrate overheating above the liner creep limit), burn-through (complete penetration of the liner wall creating an uncontrolled bypass path for hot combustion gas), dilution hole distortion (deformation of the circumferential dilution air film holes that control combustor temperature profile), and impingement cooling array blockage (fouling of the backside impingement cooling film holes from combustion product deposit). The combustor liner AI processes rendered borescope images of the combustor annulus — typically obtained by inserting through the fuel injector port access panels — and classifies each identified hot spot by area (cm²) and estimated temperature excess above nominal from the colorimetric TBC damage signature. The OEM CMM (Component Maintenance Manual) defines the serviceable hot spot limits — typically <25cm² per panel and no burn-through — and the AI classification drives the panel disposition.

Combustor liner burn-through, if not detected at the scheduled borescope inspection, allows hot combustion gas ingestion into the engine case structure — damaging the HPT outer air seal, the HPT shroud assembly, and the NGV (nozzle guide vane) support structure. GE Aviation LEAP-1B combustor liner hot section issues (EEAP — Engine Experience Action Plan, CFM International SB LEAP-1B-72-00133, 2022) required accelerated borescope inspection intervals at 1,500 EFH (equivalent flight hours) for certain serial number ranges — demonstrating that combustor liner AI classification underpinning the normal 3,000 EFH inspection interval is operationally load-bearing. An adversarially suppressed combustor liner hot spot classification that masks a 30cm² burn-through in progress as a nominal 8cm² serviceable hot spot prevents the accelerated inspection flag and the shop visit that would repair the liner before burn-through propagates to HPT damage.

4. Engine life management and CARE (Cyclic life Accumulation) AI

Life-limited parts (LLPs) in turbofan engines — the fan disk, HPC disk stack, HPT disk, and LPT disk — have certified cycle life limits established by the OEM under FAA 14 CFR 33.70 and EASA CS-E 515 “Engine Life Limitation” requirements: each LLP must be removed from service before reaching the certified cycle count regardless of apparent condition. Engine life management AI — including Rolls-Royce IntelliEngine’s CARE (Cyclic life Accumulation and Retirement Engine) algorithm, GE Aviation’s Engine Life Management system, and Pratt & Whitney’s TALON combustion system monitoring — processes borescope inspection images alongside Engine Health Monitoring (EHM) telemetry (EGT margin, oil consumption, vibration signature) to generate remaining useful life (RUL) estimates for each disk and blade population. The RUL estimate feeds the MRO’s work scope planning engine that determines whether the scheduled shop visit will include LLP replacements (high-cost: $500k–$2M per engine visit) or borescope-supported on-condition continuation. An adversarial perturbation on borescope images used by the life management AI — suppressing blade creep deformation signatures or coating degradation patterns that the AI uses as inputs to the RUL regression model — can cause the AI to overestimate remaining blade life and defer a shop visit that should be accelerated, leaving an engine population at elevated IFSD risk until the life management error is identified at the next borescope or by EHM trend exceedance.

Integration: jet engine borescope AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for borescope inspection AI belongs at the rendered endoscope image ingestion boundary before each AI classification step — before the HPT blade crack/oxidation classifier, the FOD damage severity classifier, the combustor liner hot spot classifier, and the life management AI borescope image input. Threshold 40 for borescope inspection AI reflects catastrophic failure consequence (IFSD, uncontained blade failure, hull loss) with the mitigating presence of a qualified MRO inspector review layer that validates AI classifications above internal confidence thresholds under EASA AMC 20-16 human oversight requirements.

import asyncio, base64, hashlib, json
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Borescope inspection AI: threshold 40
# FAA 14 CFR Part 43 Appendix D, EASA Part 145, AMC 20-16 MRO AI tool criteria.
BORESCOPE_AI_THRESHOLD = 40


class BorescopeAIContext(Enum):
    HPT_BLADE_CRACK         = "hpt_blade_crack"         # HPT blade crack / TBC spallation AI
    FOD_DAMAGE              = "fod_damage"               # Fan / compressor FOD damage severity AI
    COMBUSTOR_HOT_SPOT      = "combustor_hot_spot"       # Combustor liner hot spot / burn-through AI
    LIFE_MANAGEMENT         = "life_management"          # Engine life management borescope AI input


class AdversarialBorescopeImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in a borescope
    inspection AI image above threshold 40.

    Consequence if not raised: defect masked from AI classifier → continued
    service recommendation for a blade or liner requiring shop removal →
    IFSD risk or uncontained engine failure.
    Fail-safe: suppress AI classification, route image to qualified MRO
    inspector for manual dimensional assessment under EASA AMC 20-16.
    """

    def __init__(self, scan_id: str, score: int,
                 context: BorescopeAIContext,
                 engine_serial: str, blade_position: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.engine_serial = engine_serial
        self.blade_position = blade_position
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial borescope image: "
            f"context={context.value} score={score} "
            f"engine={engine_serial} blade={blade_position} scan_id={scan_id}"
        )


async def scan_borescope_image(
    image_bytes: bytes,
    context: BorescopeAIContext,
    engine_serial: str,
    engine_model: str,
    blade_position: str,
    inspection_stage: str,
    work_order_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a jet engine borescope inspection image for adversarial content.

    Fail-safe contract: AdversarialBorescopeImageError or httpx error →
    suppress AI classification, route to qualified MRO inspector for
    manual assessment. Engine must not be returned to service without
    human disposition of the affected image and blade position.

    Args:
        image_bytes: Rendered borescope endoscope frame (JPG/PNG).
        context: BorescopeAIContext identifying the inspection type.
        engine_serial: Engine serial number (e.g., 'CFM-LEAP1B-450234').
        engine_model: Engine model designation (e.g., 'LEAP-1B27').
        blade_position: Blade number / stage position (e.g., 'HPT-Stage1-Blade7').
        inspection_stage: Borescope inspection stage identifier.
        work_order_id: MRO work order number for traceability.
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict.

    Raises:
        AdversarialBorescopeImageError: if score exceeds threshold 40.
        httpx.HTTPStatusError: on Glyphward API error (fail-closed).
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"borescope:{context.value}:{engine_serial}:{blade_position}",
        "metadata": {
            "engine_serial": engine_serial,
            "engine_model": engine_model,
            "blade_position": blade_position,
            "inspection_stage": inspection_stage,
            "work_order_id": work_order_id,
            "image_sha256": image_hash,
            "context": context.value,
        },
    }
    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=5.0,  # MRO inspection is not real-time; allow generous timeout
    )
    resp.raise_for_status()
    result = resp.json()

    await _write_borescope_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        engine_serial=engine_serial,
        blade_position=blade_position,
        work_order_id=work_order_id,
        flagged=result["score"] > BORESCOPE_AI_THRESHOLD,
    )

    if result["score"] > BORESCOPE_AI_THRESHOLD:
        raise AdversarialBorescopeImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            engine_serial=engine_serial,
            blade_position=blade_position,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def _write_borescope_scan_audit(
    *, image_hash: str, scan_id: str, score: int,
    context: BorescopeAIContext, engine_serial: str,
    blade_position: str, work_order_id: str, flagged: bool,
) -> None:
    record = {
        "ts": datetime.now(timezone.utc).isoformat(),
        "scan_id": scan_id,
        "image_sha256": image_hash,
        "context": context.value,
        "score": score,
        "threshold": BORESCOPE_AI_THRESHOLD,
        "flagged": flagged,
        "engine_serial": engine_serial,
        "blade_position": blade_position,
        "work_order_id": work_order_id,
        "regulatory_refs": [
            "FAA 14 CFR Part 43 Appendix D (borescope inspection requirements)",
            "FAA 14 CFR Part 33.70 (engine life-limited part requirements)",
            "FAA Order 8900.1 Vol.6 Ch.9 Sec.12 (air carrier maintenance surveillance)",
            "EASA Part 145 (approved maintenance organisation requirements)",
            "EASA AMC 20-16 (AI tool criteria for approved maintenance)",
            "EASA CS-E 515 (engine life limitation certification)",
            "CFM International EMM — CFM56 / LEAP-1A/1B approved borescope limits",
            "NTSB Safety Recommendation A-10-115 (engine inspection AI validation)",
        ],
    }
    audit_path = Path("/var/log/glyphward/borescope_ai_scan_audit.jsonl")
    audit_path.parent.mkdir(parents=True, exist_ok=True)
    with audit_path.open("a") as fh:
        fh.write(json.dumps(record) + "\n")

Deploy scan_borescope_image at the borescope image ingestion boundary in your MRO inspection platform before each AI classification call. On AdversarialBorescopeImageError: suppress AI classification, log the flagged image and blade position to the work order, and route to a qualified MRO inspector (EASA Part 66 Cat B1 or FAA A&P certificate holder) for manual dimensional assessment against the OEM Engine Maintenance Manual serviceable limits. The engine must not receive an Aircraft Maintenance Release until the flagged blade position has been manually dispositioned. Get early access

Related questions

Further reading

• Prompt injection in aerospace MRO AI — aircraft maintenance inspection and structural health monitoring adversarial attacks
• Prompt injection in aerospace and defence AI — avionics, ISR, and autonomous systems adversarial attacks
• Prompt injection in manufacturing quality inspection AI — industrial vision inspection adversarial attacks
• Prompt injection in railway signalling AI — ETCS/ERTMS signal recognition AI bypass and PTC injection
• Prompt injection scanning API free tier — 10 scans/day, no card required