Prompt injection in aerospace MRO AI

Four adversarial injection surfaces in aerospace MRO AI

1. Aircraft component inspection photograph AI injection (Airbus Skywise AI, GE Aviation AI, AFI KLM E&M AI)

Aircraft component inspection AI processes non-destructive testing (NDT) and non-destructive inspection (NDI) photographs, component condition images, structural inspection frame images, and fatigue crack survey photographs submitted through AI-assisted maintenance management platforms that extract defect classifications — crack indication, corrosion severity, fatigue damage, wear measurement — from these image inputs and generate maintenance action recommendations that determine whether an aircraft component is approved for continued service, mandated for repair, or classified as unserviceable and subject to immediate removal. Airbus Skywise AI — the predictive maintenance and fleet health management platform operating under data sharing agreements with major Airbus operators including Air France, Delta Air Lines, Lufthansa Technik, Singapore Airlines, and Qatar Airways — processes structural inspection and NDT photograph submissions from line maintenance and base MRO operations through AI-assisted defect detection and airworthiness evaluation tools that generate component serviceability classifications governing dispatch and airworthiness directive compliance. GE Aviation Digital Solutions AI processes engine component inspection photographs — including compressor blade leading edge condition images, turbine blade tip clearance photographs, hot section combustor liner inspection images, and LPT (low pressure turbine) stage condition photographs — through AI engine health management tools that generate engine shop visit recommendations and on-wing inspection action items for the 25,000+ engines under active GE monitoring contracts with commercial airlines. AFI KLM E&M AI processes aircraft component inspection photographs across its heavy maintenance and line maintenance operations at Amsterdam Schiphol and Toulouse, generating component serviceability records that form part of the aircraft maintenance release documentation required by EASA Part 145 and the relevant national aviation authority before the aircraft is returned to service.

The adversarial injection surface is the NDT/NDI inspection photograph and component condition image submission pathway: photographs of aircraft structural components, engine parts, landing gear assemblies, and avionics bay inspections submitted to Airbus Skywise AI, GE Aviation AI, or AFI KLM E&M AI for AI defect classification and maintenance action generation. An adversarially crafted component inspection photograph — in which pixel perturbations applied to the crack indication region, corrosion surface area, or fatigue crack tip of a structural component NDT image cause the AI to classify the defect as below the serviceable limit when the unperturbed image would trigger an unserviceable classification — can suppress a maintenance action that would otherwise ground the aircraft for repair, allowing a component with an undetected structural defect to remain in service. The safety consequence of adversarial suppression of an airworthiness defect flag in an inspection AI is that the aircraft continues in revenue service with a component that has not received the mandated maintenance action — this is the canonical pathway by which undetected aircraft structural defects progress to in-service failure.

The regulatory consequences of adversarially suppressed component defect detection in aerospace inspection AI are among the most severe of any industry sector. FAA Advisory Circular AC 43.13-1B (Acceptable Methods, Techniques, and Practices — Aircraft Inspection and Repair) defines the standards for NDT inspection methods and the criteria for classifying inspection findings as within-limits or beyond-limits for continued airworthiness; failure to perform required maintenance actions identified through inspection creates airworthiness certificate violations under 14 CFR Part 91. EASA Part 145 (Approved Maintenance Organisation) requirements impose quality management system obligations on MRO organisations that include verification of inspection findings and maintenance action records; an AI-generated inspection finding that was adversarially manipulated to suppress a defect classification creates a Part 145 quality system failure that can result in EASA AMO approval suspension. Airworthiness Directive (AD) compliance — mandatory maintenance actions mandated by FAA or EASA in response to fleet safety findings — is legally required under 14 CFR Part 39; an aircraft that cannot demonstrate AD compliance faces immediate grounding and certificate action. 18 USC § 32 (destruction of aircraft or aircraft facilities) creates criminal liability for any person who knowingly performs an act that would render an aircraft unairworthy, with sentences of up to 20 years imprisonment for acts that endanger life; the adversarial suppression of an inspection AI defect flag that allows a structurally deficient component to remain in service represents a potential violation of the constructive airworthiness destruction standard. Threshold: 50 for aircraft component inspection AI — the strictest Glyphward threshold, reflecting aviation safety primacy.

2. Engine borescope and FOQA image AI injection (GE Aviation AI, Safran Aircraft Engines AI, Rolls-Royce FAST AI)

Engine borescope and flight operational quality assurance (FOQA) AI processes engine borescope inspection photographs, endoscope images of hot section components, FOQA data display screenshots, and engine test cell result photographs submitted through AI-assisted engine health management systems that extract damage classification, hot section condition assessment, and engine performance parameter data from these image inputs to generate engine shop visit recommendations, on-wing inspection action items, and power-by-the-hour (PbH) contract cost avoidance reports for commercial airline and lessor clients. GE Aviation Digital Solutions AI processes borescope inspection photographs from on-wing engine inspections of GEnx, GE9X, GE90, and CF6 engines across its 25,000+ engine monitoring contract fleet, extracting hot section damage severity classifications — combustor liner burnout, HPT blade tip curl, NGV oxidation, LPT blade erosion — from borescope images submitted by airline line maintenance technicians through GE’s On Point engine monitoring portal. Safran Aircraft Engines AI processes borescope inspection photographs for CFM International LEAP-1A and LEAP-1B engines in service on the A320neo and 737 MAX fleets, extracting compressor and turbine stage condition data from on-wing borescope submissions to generate LEAP engine health recommendations distributed to CFM’s airline customers and MRO partners. Rolls-Royce FAST (Fleet Management, Asset Management, Support and Training) AI processes borescope and engine health monitoring images for Trent 700, Trent 1000, Trent XWB, and Trent 7000 engines, generating shop visit recommendations and on-wing action items for the Rolls-Royce TotalCare PbH contract fleet under which engine shop visits are scheduled and cost-shared between Rolls-Royce and its airline customers.

The adversarial injection surface is the engine borescope photograph and FOQA data display screenshot submission pathway: borescope images of engine hot section components, endoscope photographs of compressor and turbine stages, FOQA parameter display screenshots, and engine test cell result photographs submitted by airline maintenance technicians, MRO facility engineers, or third-party borescope service providers through GE Aviation, Safran, or Rolls-Royce AI engine health management portals for AI damage classification and shop visit recommendation generation. An adversarially crafted borescope image — in which pixel perturbations applied to the HPT blade tip curl, combustor liner burnout indicator, or NGV oxidation region of an engine hot section borescope photograph cause the GE Aviation AI or Rolls-Royce FAST AI to classify the damage as below the shop visit trigger threshold when the unperturbed image would generate an immediate shop visit recommendation — can defer a required engine shop visit, allowing an engine with undetected hot section damage to continue in revenue service while the operator avoids the $2-5M direct cost of an unplanned engine removal and shop visit. In PbH (Power by the Hour) contract contexts, where Rolls-Royce, CFM, or GE absorbs the cost of scheduled shop visits under agreed cost-per-engine-flight-hour rates, adversarial borescope injection that defers a shop visit by a single scheduled interval fraudulently reduces the cost exposure borne by the engine OEM under the PbH contract while increasing the operator’s undetected maintenance liability.

The regulatory and commercial consequences of adversarially manipulated engine borescope AI classifications are severe across safety, criminal, and contract dimensions. FAA Technical Standard Order (TSO) authorisation requirements for engine inspection procedures and the FAA’s Type Certificate Data Sheet (TCDS) compliance obligations for specific engine models impose mandatory inspection intervals and damage classification criteria that must be met for the engine to remain airworthy under its Type Certificate — an AI-generated borescope classification that was adversarially suppressed to fall below the TCDS damage threshold creates an airworthiness violation under 14 CFR Part 33 (Airworthiness Standards: Aircraft Engines). 18 USC § 1001 (false statements and concealment in federal matters) applies to false maintenance records submitted to the FAA, including maintenance records that incorporate adversarially manipulated AI-generated engine inspection findings as the basis for airworthiness release documentation — criminal liability attaches to each false entry made with knowledge of its falsity. Power by the Hour contract fraud through adversarially manipulated engine inspection AI creates civil contract liability for the airline operator and potentially criminal wire fraud exposure under 18 USC § 1343 where the adversarially manipulated inspection data is transmitted across state lines to the engine OEM’s monitoring systems. Threshold: 50 for engine borescope and FOQA image AI.

3. Aircraft maintenance records document scan AI injection (Boeing Analytix AI, MRO Pro AI, OASES AI)

Aircraft maintenance records AI processes scanned maintenance logbook entries, airworthiness release document scans, AD compliance record photographs, component traceability document images, and engine build record scans submitted through AI-assisted maintenance management and records compliance platforms that extract maintenance event dates, component part numbers, AD reference numbers, and certification statement data from these document image inputs, generating maintenance compliance histories and airworthiness documentation records that form the legal basis for the aircraft’s Certificate of Airworthiness status and for commercial aircraft transactions including sales, lease placements, and financing arrangements. Boeing Analytics AI (Boeing Analytix) processes maintenance record scans and fleet documentation photographs from Boeing commercial aircraft operator submissions, extracting maintenance event and airworthiness compliance data through AI-assisted document processing tools integrated into Boeing’s fleet support and maintenance management services for the 737, 767, 777, and 787 programme fleets. MRO Pro AI processes aircraft maintenance record document scans through AI-assisted records management tools deployed at commercial MRO providers, extracting part number traceability data, maintenance release signatures, and AD compliance status from scanned technical log entries, work package records, and release-to-service documents. OASES (Operator Aircraft and Servicing Enterprise Suite) AI, deployed by 150+ airline and MRO operator clients worldwide including TUI Airways, Air Arabia, and Ryanair technical operations, processes maintenance record document submissions through AI-assisted airworthiness management and records compliance tools that extract maintenance release and AD compliance data for integration into the OASES maintenance programme management system. Ramco Aviation AI, deployed for fleet maintenance management at airline and MRO operator clients including Air India, Air Arabia, and JAL, processes maintenance record document images through AI-assisted task card completion tracking and airworthiness release documentation workflows.

The adversarial injection surface is the maintenance logbook scan, airworthiness release document photograph, and AD compliance record image submission pathway: scanned paper maintenance logbook entries, photographs of manual work package completion records, digitised AD compliance sign-off sheets, and component traceability document scans submitted to Boeing Analytix AI, MRO Pro AI, OASES AI, or Ramco Aviation AI for AI data extraction and maintenance compliance record generation. An adversarially crafted maintenance logbook scan or AD compliance record photograph — in which pixel perturbations applied to the date field, part number region, certifying engineer licence number, or AD reference number area of a scanned maintenance document cause the Boeing Analytix AI or OASES AI to extract false compliance data — can generate a fraudulent maintenance compliance record that shows an AD as complied-with when the actual physical maintenance task was not performed, or records a component part number that falsely passes a life-limit traceability check when the actual installed component has exceeded its certified life limit. The falsified maintenance compliance record generated from an adversarially manipulated document scan enters the aircraft maintenance management system as the authoritative compliance record for the affected maintenance event, where it governs airworthiness release decisions and is incorporated into the aircraft’s permanent technical records.

The regulatory consequences of adversarially falsified maintenance record data extracted by AI from manipulated document scans are severe under both administrative and criminal aviation law. FAA 14 CFR Part 43 § 43.9 (Content, form, and disposition of maintenance records) requires that aircraft maintenance records accurately document the maintenance performed, the date of performance, the identity of the person performing the maintenance, and the approved data used — maintenance records that incorporate adversarially falsified AI-extracted data are false maintenance records under § 43.9, with the falsification constituting a violation of the Certificate of Airworthiness maintenance requirements. 18 USC § 1001 (false statements in federal matters) applies to false entries in aircraft maintenance records that are submitted to or incorporated into FAA-regulated airworthiness documentation, with five-year imprisonment penalties for knowing falsification. AD non-compliance discovered during FAA ramp inspection or audit can result in immediate aircraft grounding, civil penalty of up to $1,100 per day of continuing violation under 14 CFR Part 13, and referral for criminal enforcement where knowing falsification of AD compliance records is identified. In commercial aircraft transactions — sale, lease placement, engine sale-leaseback, or JOLCO financing — aircraft maintenance record accuracy is a condition precedent to the transaction and is warranted by the seller or lessor; maintenance records that incorporate adversarially falsified AI data create warranty breach liability and transaction rescission risk under aviation asset purchase and lease agreement terms. Threshold: 50 for aircraft maintenance records document scan AI.

4. Aircraft storage and asset valuation AI injection (ISTAT appraisers, AVAC AI, mba Aviation AI)

Aircraft storage and asset valuation AI processes photographs of parked and stored aircraft, thermal inspection images of stored aircraft systems, maintenance status document scans, and aircraft condition survey photographs submitted through AI-assisted aviation asset management and appraisal platforms that extract aircraft storage condition classification, structural integrity assessment, systems condition data, and maintenance recency status from these image inputs, generating aircraft condition assessments and appraised base values that determine financing covenant compliance, lease return condition acceptance, and aviation asset portfolio valuations for airline lessors, aircraft financing institutions, and aviation investment funds. ISTAT (International Society of Transport Aircraft Trading) certified appraisers use AI-assisted condition assessment tools to generate aircraft base values and current market values for secured financing transactions including JOLCOs (Japanese Operating Lease with Call Option) and ECA (Export Credit Agency) financing arrangements, with AI condition assessment photographs forming part of the appraisal evidence record supporting the appraised value. AVAC (Aircraft Value Analysis Co.) AI processes aircraft condition survey photographs and stored aircraft inspection images through AI-assisted valuation tools that generate aircraft appraised values for aircraft lessors, airline operators, and aviation investors. mba Aviation AI processes aircraft and engine asset condition photographs through AI-assisted appraisal and condition assessment tools for commercial aircraft transaction advisory and portfolio management services. Collateral Management International (CMI) AI processes aircraft condition survey photographs for aviation asset finance and operating lease portfolios, generating condition assessment reports that determine lease return acceptance and maintenance reserve credit adjustments for airline and lessor clients.

The adversarial injection surface is the stored aircraft condition photograph, thermal inspection image, and aircraft maintenance status document scan submission pathway: photographs of parked aircraft at storage facilities (Victorville VCVAFD, Tucson DMAFB, Alice Springs, Teruel), thermal images of stored aircraft control surfaces and hydraulic systems, maintenance status document scans showing compliance with storage maintenance programme requirements, and aircraft condition survey photographs submitted to ISTAT appraiser AI tools, AVAC AI, mba Aviation AI, or CMI AI for AI condition assessment and appraised value generation. An adversarially crafted aircraft storage condition photograph — in which pixel perturbations applied to the fuselage skin surface image, control surface condition region, or landing gear strut area of a stored aircraft photograph cause the AI to classify the aircraft storage condition as airworthy-ready-for-return-to-service when the unperturbed photograph would indicate significant storage deterioration, corrosion, or systems degradation — can generate an inflated aircraft condition assessment and appraised base value that supports a higher financing loan-to-value ratio, a more favourable lease return condition acceptance, or a fraudulently elevated aviation asset portfolio valuation. In JOLCO financing structures, where the appraised aircraft value determines the loan principal and the tax benefit calculation for the Japanese investor partner, adversarially inflated AI condition assessment inputs can create fraudulent financing valuations with multi-million dollar financial exposure for aviation banks and ECA guarantee providers.

The regulatory and contractual consequences of adversarially manipulated aircraft storage condition AI classifications span aviation finance, securities, and criminal law dimensions. ISTAT Appraiser Standards impose professional liability obligations on certified appraisers whose AI-assisted appraisal tools generate values based on adversarially manipulated condition assessment inputs — an appraised value that was inflated through adversarial image injection into the appraiser’s AI tool creates professional negligence exposure for the appraiser and misrepresentation liability for the financing party that relied on the inflated value. JOLCO and ECA financing documentation — including loan agreements, security deposit agreements, and ECA guarantee applications — incorporates maintenance covenants and condition representation warranties based on appraised aircraft values; a financing covenant breach caused by adversarially inflated AI condition assessment data creates event-of-default risk under the loan agreement and potential ECA guarantee claim disputes. Aircraft lease return condition acceptance processes governed by lease agreement technical provisions use AI condition assessment outputs as inputs to the maintenance reserve credit and return condition adjustment calculations — adversarially manipulated condition photographs that inflate AI return condition scores reduce maintenance reserve credits payable by the airline lessee, creating unjust enrichment at the lessor’s expense. Threshold: 55 for aircraft storage and asset valuation AI, reflecting the financial fraud dimensions above the pure safety context.

Integration: aerospace MRO AI image ingestion with Glyphward pre-scan

Aerospace MRO AI image ingestion flows from aircraft maintenance facility NDT portals, borescope image upload interfaces, maintenance record scan workflows, and aircraft condition survey photograph channels into inspection AI, engine health AI, records compliance AI, and valuation AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary — particularly for externally sourced inspection photographs, third-party borescope submissions, maintenance record document scans, and aircraft condition survey images — before the AI-generated output is committed to maintenance records, airworthiness releases, or financing valuations:

import asyncio
import base64
import hashlib
import os
import uuid
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = os.environ["GLYPHWARD_API_KEY"]
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Aerospace MRO AI — suppression of airworthiness defect flags, engine hot
# section damage concealment, maintenance record falsification, aircraft
# storage condition inflation for financing fraud.
# FAA 14 CFR Part 43, EASA Part 145, 18 USC §32 (aircraft sabotage),
# 18 USC §1001 (false statements to FAA), FAA ADs, ISTAT Appraiser Standards.
THRESHOLD_MRO_SAFETY    = 50   # component inspection, borescope, maintenance records
THRESHOLD_MRO_VALUATION = 55   # aircraft storage and asset valuation


class AerospaceMROAIContext(str, Enum):
    COMPONENT_INSPECTION = "component_inspection"  # Airbus Skywise, GE Aviation, AFI KLM E&M
    ENGINE_BORESCOPE     = "engine_borescope"      # GE Aviation, Safran, Rolls-Royce FAST
    MAINTENANCE_RECORDS  = "maintenance_records"   # Boeing Analytix, MRO Pro, OASES
    ASSET_VALUATION      = "asset_valuation"       # ISTAT appraisers, AVAC, mba Aviation


def threshold_for(context: AerospaceMROAIContext) -> int:
    if context == AerospaceMROAIContext.ASSET_VALUATION:
        return THRESHOLD_MRO_VALUATION
    return THRESHOLD_MRO_SAFETY


async def scan_mro_image(
    image_path: str | Path,
    context: AerospaceMROAIContext,
    plant_id_hash: str,     # SHA-256 of MRO facility / airline identifier
    aircraft_hash: str,     # SHA-256 of aircraft registration or MSN
    work_order_ref: str,    # e.g. "WO-2026-44721", "AD-2026-10-12", "JOLCO-JA835A"
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan an aerospace MRO AI image for adversarial injection payloads
    before forwarding to an aircraft component inspection AI, engine borescope
    AI, maintenance records AI, or aircraft storage valuation AI.

    Raises AdversarialMROImageError if the Glyphward score meets or exceeds
    the context-specific threshold (50 for safety contexts, 55 for valuation).

    Safety contexts use threshold 50 — the strictest Glyphward setting:
      - COMPONENT_INSPECTION:  FAA 14 CFR Part 43, EASA Part 145,
                               18 USC §32 (aircraft sabotage), FAA ADs
      - ENGINE_BORESCOPE:      FAA TCDS Part 33, 18 USC §1001 (false
                               maintenance records), PbH contract fraud
      - MAINTENANCE_RECORDS:   14 CFR §43.9 false records, 18 USC §1001,
                               AD non-compliance $1,100/day civil penalty

    Valuation context uses threshold 55:
      - ASSET_VALUATION:       ISTAT Appraiser Standards professional
                               liability, JOLCO/ECA financing fraud,
                               lease return condition misrepresentation
    """
    image_bytes = Path(image_path).read_bytes()
    image_b64   = base64.b64encode(image_bytes).decode()
    image_sha256 = hashlib.sha256(image_bytes).hexdigest()
    client_scan_id = str(uuid.uuid4())
    threshold = threshold_for(context)

    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json={
            "image": image_b64,
            "source": context.value,
            "metadata": {
                "mro_context":      context.value,
                "plant_id_hash":    plant_id_hash,
                "aircraft_hash":    aircraft_hash,
                "work_order_ref":   work_order_ref,
                "client_scan_id":   client_scan_id,
                "image_sha256":     image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "plant_id_hash":   plant_id_hash,
        "aircraft_hash":   aircraft_hash,
        "work_order_ref":  work_order_ref,
        "mro_context":     context.value,
        "scan_id":         result["scan_id"],
        "client_scan_id":  client_scan_id,
        "image_sha256":    image_sha256,
        "score":           result["score"],
        "flagged_region":  result.get("flagged_region"),
        "threshold":       threshold,
        "action":          "blocked" if result["score"] >= threshold else "allowed",
    }
    await write_mro_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialMROImageError(
            f"MRO AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"aircraft={aircraft_hash} ref={work_order_ref}"
        )
    return result


async def write_mro_audit_record(record: dict) -> None:
    """Persist audit record to MRO compliance audit store (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialMROImageError(Exception):
    """Raised when an aerospace MRO AI image exceeds the adversarial injection threshold."""
    pass

Call scan_mro_image() with AerospaceMROAIContext.COMPONENT_INSPECTION before forwarding NDT/NDI inspection photographs to Airbus Skywise AI, GE Aviation AI, or AFI KLM E&M AI — this is the highest safety-consequence integration point, where adversarial suppression of a structural defect flag creates direct airworthiness risk with 18 USC § 32 criminal exposure. Call with AerospaceMROAIContext.ENGINE_BORESCOPE for borescope photograph submissions before GE Aviation, Safran, or Rolls-Royce FAST AI engine health analysis, using work_order_ref to link the scan audit record to the engine shop visit recommendation chain for PbH contract and TCDS compliance audit purposes. Call with AerospaceMROAIContext.MAINTENANCE_RECORDS for maintenance logbook scans and AD compliance document photographs before Boeing Analytix AI, MRO Pro AI, or OASES AI data extraction, preserving the image_sha256 as the forensic anchor for 14 CFR § 43.9 audit trail reconstruction under FAA enforcement proceedings. Call with AerospaceMROAIContext.ASSET_VALUATION for aircraft condition survey photographs and storage assessment images before ISTAT appraiser AI tools, AVAC AI, or mba Aviation AI condition classification, with the aircraft_hash parameter linking Glyphward scan records to specific aircraft MSNs for JOLCO covenant compliance audit evidence. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four aerospace MRO AI injection surfaces; covers how adversarial pixel-level perturbations cause AI misclassification through image content manipulation without detectable visual artifacts at human review resolution.
• Vision-language model security — technical architecture of adversarial image attacks against vision-language models including the pixel perturbation classes applicable to NDT inspection photograph injection and borescope image manipulation.
• Prompt injection scanner for document AI — document AI scanning covering the broader class of scanned document injection vectors applicable to maintenance logbook scan and AD compliance record manipulation.
• Prompt injection scanner for healthcare AI — medical image injection attacks with parallel regulatory context to aerospace NDT inspection AI — both operate in safety-critical contexts with professional liability, regulatory certification, and criminal consequences for AI output falsification.
• Free tier — 10 scans/day, no card required — start scanning aerospace MRO AI images at development volumes before committing to a production plan.