Prompt injection in urban air mobility and eVTOL AI

Four adversarial injection surfaces in urban air mobility and eVTOL AI

1. Obstacle detection and collision avoidance image injection (FAA 14 CFR §91.111, §91.181, EASA CS-SC-VTOL-01 AMC)

Obstacle detection and collision avoidance AI in commercial eVTOL operations processes continuous image streams from onboard forward-facing, lateral, and downward-facing cameras capturing low-altitude urban airspace along flight corridors connecting vertiport networks in metropolitan areas including the Los Angeles Basin where Archer Aviation Midnight is targeting 2025–2026 commercial launch, the San Francisco Bay Area where Joby Aviation is targeting initial commercial service under FAA Part 135 air carrier certification in progress, and Miami-Dade County where Archer Aviation and Joby Aviation have identified early market launch opportunities across inter-modal urban air taxi corridor networks. The AI classification pipeline processes camera image frames to assign obstacle class labels including buildings and architectural protrusions, construction cranes and construction equipment, electrical transmission towers and power distribution lines, low-altitude manned helicopters and other aircraft on conflicting headings, communications towers and antenna arrays, advertising signage and bridge structures, and bird flocks above minimum strike-risk population thresholds — and outputs per-obstacle threat scores incorporating obstacle class confidence, estimated distance and closure rate, and collision trajectory probability that feed into autonomous or pilot-assisted collision avoidance manoeuvre execution decisions at flight altitudes between 500 and 2,500 feet AGL across urban vertiport approach corridors where FAA 14 CFR §91.181 course-to-be-flown obligations and §91.111 right-of-way and collision avoidance requirements apply to all civil aircraft. Joby Aviation’s 6-seat eVTOL with 150-plus-mile range processes camera images through detect-and-avoid AI for every flight segment; Archer Aviation Midnight processes camera obstacle detection inputs through AI-assisted detect-and-avoid functions in hybrid pilot-assisted and future autonomous configurations; Wisk Aero’s autonomous CORA 6th generation aircraft requires fully autonomous detect-and-avoid AI without pilot onboard across its Type Certification proceeding with the FAA.

The adversarial injection surface is the onboard camera image frame submission pathway into the obstacle classification and threat scoring AI pipeline. An adversarially crafted camera image frame — in which pixel perturbations imperceptible to human visual inspection are applied to the image region containing the obstacle target, specifically the pixel neighbourhood surrounding a power line span between transmission towers crossing the flight corridor, a construction crane boom at a building project within the approach path, or the fuselage and rotor disc of a low-altitude helicopter on a converging heading — can cause the obstacle classification AI to suppress the threat classification for that obstacle, reducing the per-obstacle threat score below the collision avoidance manoeuvre execution threshold when the actual camera sensor data evidences an obstacle requiring evasive action. In eVTOL operations where Joby Aviation AI or Wisk Aero AI processes camera frames at video frame rates without human pilot examination of every individual camera frame before the AI threat score governs detect-and-avoid response, adversarial suppression of obstacle threat classification creates the most operationally severe injection consequence in the eVTOL AI attack surface taxonomy: a commanded flight path through an undetected obstacle. Power lines are the highest-consequence suppression target because they are the obstacle class least visible in low-contrast sky background conditions, most frequently encountered across urban flight corridors, and most likely to cause catastrophic structural damage and in-flight loss of control on contact with rotor systems or fixed lift surfaces at eVTOL cruise velocities.

The FAA 14 CFR §91.111, §91.181, FAA Type Certificate safety case, and EASA CS-SC-VTOL-01 AMC regulatory consequences of adversarially suppressed obstacle detection classification span FAA 14 CFR §91.111 right-of-way rules establishing that powered aircraft must give way to unpowered aircraft and that no aircraft may operate so close to another aircraft as to create a collision hazard — adversarial suppression of a conflicting helicopter threat classification in Joby Aviation or Archer Aviation collision avoidance AI creates a §91.111 violation dimension and a Type Certificate safety case non-compliance dimension for any AI-assisted collision avoidance function relied upon in the FAA-accepted means of compliance for the detect-and-avoid airworthiness requirement; FAA 14 CFR §91.181 course-to-be-flown obligations establishing that each aircraft must be operated along the correct airspace route and that deviations require ATC coordination — flight path deviation commands generated by adversarially corrupted obstacle AI create §91.181 airspace management dimensions; FAA 14 CFR Part 21 Subpart H special class type certification establishing that Joby Aviation, Wisk Aero, and Supernal must demonstrate that each AI safety function meets accepted airworthiness standards through issue papers and means of compliance approved by the FAA Aircraft Certification Service — adversarial injection in obstacle detection AI creates a Type Certificate safety case validity dimension that may require issue paper revision and means of compliance re-demonstration; and EASA CS-SC-VTOL-01 AMC obstacle avoidance performance requirements applicable to Volocopter VoloCity European type certification and Joby Aviation EASA validation proceeding establishing performance-based obstacle avoidance safety objectives that adversarially corrupted detection AI may fail to meet across the required obstacle encounter scenario test matrix. Threshold: 65 for obstacle detection and collision avoidance AI — reflecting FAA §91.111 collision avoidance, §91.181 course compliance, FAA Part 21 Subpart H Type Certificate safety case, and EASA CS-SC-VTOL-01 AMC performance obligation dimensions.

2. Vertiport approach and landing zone visual AI bypass (FAA Vertiport Engineering Brief VEB-1, EASA Vertiport Design Specifications, P.L. 117-203 §4)

Vertiport approach and landing zone AI processes camera images submitted from aircraft-mounted approach cameras, fixed ground-based vertiport surveillance cameras, and sensor fusion display image overlays showing landing pad occupancy detection status, final approach surface condition assessment, wind direction sock and indicator classification, obstacle clearance zone geometry verification, and surface foreign object debris (FOD) detection from Joby Aviation AI at initial commercial vertiport facilities targeting launch in partnership with Uber Elevate and Delta Air Lines vertiport infrastructure developments; Supernal AI under Hyundai Motor Group’s S-A2 eVTOL 2028 commercial operations target at integrated ground transportation and air mobility hub vertiport infrastructure across US urban markets; and Volocopter VoloCity AI at Singapore Changi Airport vertiport facilities where CAAS regulatory certification for commercial eVTOL passenger operations was targeting 2025–2026 entry into service. The AI classification pipeline processes landing zone camera images to assign occupancy status (clear, occupied, obstructed), surface condition grade (serviceable, caution, unsafe), obstacle clearance zone status (clear, obstructed, marginal), and final approach path clearance determination (authorized, hold, go-around commanded) that govern approach continuation or go-around execution commands issued to the pilot or to the autonomous flight management system for highly automated eVTOL landing sequence execution at vertiports where air traffic management and ground handling operations rely on AI landing zone clearance determinations to maintain approach interval and departure scheduling across the vertiport network.

The adversarial injection surface is the vertiport approach camera image and landing zone surveillance display image submission pathway: Joby Aviation AI, Supernal AI, or Volocopter VoloCity AI landing zone status display images and approach camera frames submitted through AI-assisted landing clearance and approach authorization systems for AI-generated landing authorization record and go-around command generation. An adversarially crafted landing zone camera image — in which pixel perturbations applied to the landing pad surface occupancy indicator region, the obstacle clearance zone marker display, or the approach surface clearance grid cause the AI to classify a landing pad that is occupied by a ground service vehicle, obstructed by ground personnel, or bordered by a surface obstruction within the required obstacle clearance zone as a clear and authorized landing surface meeting FAA Vertiport Engineering Brief VEB-1 approach surface obstruction clearance requirements — can authorize a landing approach that the actual landing zone occupancy and obstruction status would require to be converted to a go-around command or a hold instruction, creating a collision risk between the approaching eVTOL and ground personnel, service vehicles, or surface obstacles in the obstacle clearance zone at the landing pad perimeter. In vertiport operations where Joby Aviation AI or Volocopter VoloCity AI processes landing zone camera images at approach intervals of three to five minutes without individual human landing zone controller examination of every AI-processed camera image before the AI clearance determination governs approach continuation or go-around execution, adversarial manipulation of landing zone images creates vertiport approach safety and regulatory consequences under FAA VEB-1, EASA Vertiport Design Specification, and P.L. 117-203 §4 vertiport safety standard dimensions.

The FAA Vertiport Engineering Brief VEB-1, EASA Vertiport Design Specification, P.L. 117-203 §4, and FAA 14 CFR Part 135 regulatory consequences of adversarially corrupted landing zone classification span FAA Vertiport Engineering Brief VEB-1 approach surface obstruction clearance area requirements establishing approach surface slope gradients, obstacle clearance zone geometry, and landing pad final approach and take-off (FATO) area minimum dimensions and obstruction exclusion standards that vertiport operators must comply with under FAA Advisory Circular and engineering brief authority — adversarial injection in landing zone AI that classifies an obstructed FATO area as clear creates VEB-1 approach surface compliance failure dimensions and FAA vertiport operating approval revocation exposure; EASA Vertiport Design Specification approach and departure surface obstruction clearance requirements applicable to Volocopter VoloCity Singapore certification and European vertiport operations establishing EASA performance-based obstruction clearance standards that adversarially corrupted approach AI may fail across the required approach scenario evaluation matrix; P.L. 117-203 Advanced Air Mobility Coordination and Leadership Act §4 vertiport safety standard development mandate establishing Congressional direction for FAA vertiport safety standard promulgation that includes landing zone AI classification reliability requirements within the scope of vertiport operational safety standards being developed for the commercial AAM network; and FAA 14 CFR Part 135 air carrier operation specifications applicable to Joby Aviation and Archer Aviation on-demand passenger operations establishing that operations specifications must include approved landing area and vertiport qualification standards — adversarial injection in landing zone AI that bypasses approach obstruction detection creates Part 135 operations specifications compliance dimensions. Threshold: 70 for vertiport approach and landing zone AI — reflecting FAA VEB-1 approach surface obstruction clearance, EASA Vertiport Design Specification performance standards, P.L. 117-203 §4 safety mandate, and FAA Part 135 operations specifications dimensions.

3. Passenger biometric boarding and identity verification bypass (TSA Identity-Based Boarding, FAA 14 CFR §135.63, EU AI Act Art. 5)

Passenger biometric boarding and identity verification AI processes facial recognition camera images captured at Joby Aviation and Archer Aviation passenger terminal boarding lanes to verify passenger identity against flight reservation manifest, replace conventional boarding pass document presentation with biometric boarding authorization, confirm passenger weight and identity for FAA 14 CFR §135.63 weight and balance manifest documentation applicable to on-demand air carrier operations, and generate boarding completion records that form the basis for departure authorization and passenger manifest certification required by FAA Part 135 operations specifications. The AI classification pipeline processes terminal boarding camera images to output passenger identity confidence scores against enrolled reservation biometric templates, identity verification approval or denial determinations, weight manifest assignment confirmations, and boarding authorization records that govern gate release for eVTOL departure across Joby Aviation’s targeted commercial launch cities and Archer Aviation Midnight’s Los Angeles and Miami commercial service network. Biometric boarding AI is also subject to EU AI Act Article 5(1)(d) real-time remote biometric identification system prohibition in publicly accessible spaces applicable to Volocopter VoloCity and Archer Aviation European market operations, creating a dual regulatory framework where US TSA Identity-Based Boarding programme requirements and EU AI Act Article 5 prohibition boundaries must both be navigated in cross-jurisdictional eVTOL boarding AI deployments.

The adversarial injection surface is the passenger boarding camera image submission pathway: Joby Aviation AI or Archer Aviation AI terminal boarding lane facial recognition camera images submitted through AI-assisted passenger identity verification and weight manifest documentation tools for AI-generated boarding authorization record creation. An adversarially crafted facial image — in which pixel perturbations applied to the passenger facial region captured by the boarding lane camera, to the identity document facial photograph region on a government-issued ID document image, or to the facial comparison result display image showing match confidence score and identity verification status, cause the AI to classify an unauthorized passenger whose facial biometric does not match any enrolled reservation as a verified match for a specific reservation, or to match a passenger to a different reservation with a different weight manifest assignment — can allow unauthorized persons to board eVTOL flights at Joby Aviation or Archer Aviation vertiport terminals, cause passenger weight manifest documentation required under §135.63 to record incorrect passenger identity and weight assignments affecting weight and balance calculations for the departing aircraft, or generate a false boarding completion record that certifies passenger identity verification was performed when the biometric match was adversarially corrupted. In eVTOL terminal boarding operations where Joby Aviation AI or Archer Aviation AI processes boarding camera images at departure intervals without continuous human identity verification agent oversight of every AI-processed boarding image before the AI authorization governs gate release and manifest certification, adversarial manipulation of boarding camera images creates FAA §135.63 manifest fraud and EU AI Act Art. 5 regulatory dimensions.

The FAA 14 CFR §135.63, TSA Identity-Based Boarding, EU AI Act Art. 5(1)(d), and aviation security regulatory consequences of adversarially corrupted boarding biometric classification span FAA 14 CFR §135.63 passenger manifest requirements establishing that Part 135 air carrier operators must prepare accurate weight and balance manifests for each flight containing passenger identity and weight information — adversarial corruption of boarding biometric AI that assigns a passenger to an incorrect reservation creates §135.63 manifest inaccuracy dimensions with FAA Part 135 certification compliance consequences and potential Certificate Action implications for repeated manifest documentation failures; TSA Identity-Based Boarding programme framework establishing Transportation Security Administration requirements for biometric boarding technology deployment at US airport and vertiport facilities including system accuracy, liveness detection, and watchlist comparison obligations that adversarially corrupted biometric boarding AI may fail to meet across the required performance evaluation matrix; EU AI Act Article 5(1)(d) prohibition on real-time remote biometric identification systems in publicly accessible spaces applicable to eVTOL vertiport terminal facial recognition boarding AI deployed at publicly accessible terminal facilities in EU member states — adversarially circumvented biometric boarding AI that performs real-time identification of persons not enrolled in the boarding system creates prohibited-use classification dimensions under EU AI Act Art. 5 enforcement by data protection authorities; and ICAO Annex 9 Facilitation standards applicable to international eVTOL passenger operations establishing passenger documentation and identity verification requirements for transborder air carrier operations. Threshold: 60 for passenger biometric boarding AI — reflecting FAA 14 CFR §135.63 manifest documentation, TSA Identity-Based Boarding compliance, EU AI Act Art. 5(1)(d) biometric prohibition, and ICAO Annex 9 Facilitation standard dimensions.

4. Advanced air traffic management display image injection (FAA UTM ConOps, EASA U-space EU 2021/664–666, ICAO Annex 11)

Advanced air traffic management display AI in urban air mobility operations processes airspace display images generated by UTM service suppliers and U-space service providers including NASA UTM AI, AirMap AI, and Altitude Angel AI that show traffic conflict indicator overlays for aircraft on conflicting trajectories within eVTOL flight corridor networks, geo-fence boundary compliance display images showing eVTOL aircraft position relative to approved operating area boundaries and restricted airspace perimeters, NOTAM (Notice to Air Missions) map overlay images showing temporary flight restriction boundaries, airspace status notifications, and obstacle advisories affecting the eVTOL network operating area, strategic and tactical conflict detection display images showing four-dimensional trajectory conflict predictions and resolution advisory recommendations for eVTOL fleet sequencing across vertiport network approach and departure corridors, and conformance monitoring display images showing individual aircraft adherence to approved four-dimensional trajectory flight plans required under EASA U-space regulation EU 2021/666 flight authorisation and conformance monitoring service obligations. These display images are processed by flight planning AI, air traffic flow management AI, and autonomous flight management AI across the UTM service provider and U-space service provider infrastructure layers that Joby Aviation, Archer Aviation, Wisk Aero, and Supernal will rely upon for airspace integration and deconfliction during commercial eVTOL network operations.

The adversarial injection surface is the UTM traffic display image and airspace compliance status display image submission pathway: NASA UTM AI, AirMap AI, or Altitude Angel AI airspace conflict indicator display images and geo-fence compliance screens submitted through AI-assisted flight plan management and autonomous flight authorization tools for AI-generated airspace deconfliction command and flight authorization record generation. An adversarially crafted UTM traffic display image — in which pixel perturbations applied to the traffic conflict indicator icon for an aircraft on a converging trajectory toward a Joby Aviation or Wisk Aero eVTOL, the geo-fence boundary compliance alert indicator for an aircraft approaching a restricted airspace boundary, or the NOTAM boundary overlay display for an active temporary flight restriction cause the UTM AI to suppress the conflict indicator and classify the airspace display as conflict-free and clear for continued operations when the actual aircraft position and trajectory data evidences an unresolved traffic conflict requiring immediate tactical conflict resolution advisory — can suppress a traffic conflict indicator that would otherwise generate an ATC coordination requirement, a tactical manoeuvre advisory to the conflicting aircraft, or a strategic re-routing command that would resolve the conflict before it reaches minimum separation standards. In UTM operations where NASA UTM AI or AirMap AI processes airspace display images for conflict detection across tens to hundreds of simultaneously operating eVTOL and drone aircraft without human air traffic controller examination of every AI-processed display image before the AI conflict assessment governs tactical resolution advisory issuance, adversarial suppression of traffic conflict indicators creates ICAO Annex 11 Air Traffic Services minimum separation and conflict resolution dimensions.

The FAA UTM ConOps, EASA U-space EU 2021/664–666, ICAO Annex 11, and Advanced Air Mobility Coordination and Leadership Act regulatory consequences of adversarially suppressed UTM conflict classification span FAA UAS Traffic Management ConOps establishing operational concept requirements for UTM service supplier accuracy and conflict detection performance at the USS provider service level applicable to AirMap AI and other FAA-recognised UTM service suppliers operating in FAA designated UTM corridors where Joby Aviation and Wisk Aero conduct eVTOL operations — adversarial suppression of conflict indicators in UTM display AI creates USS provider service accuracy requirement failure dimensions that may affect FAA UTM participation eligibility; EASA U-space regulation EU 2021/664 establishing the regulatory framework for U-space airspace and EU 2021/666 establishing flight authorisation and conformance monitoring service requirements applicable to U-space service provider AI systems including Altitude Angel AI at European eVTOL operations — adversarially corrupted UTM conflict display AI that fails to detect geo-fence boundary violations or conflict alerts creates EU 2021/666 conformance monitoring service obligation failure dimensions with European Union Aviation Safety Agency enforcement authority; ICAO Annex 11 Air Traffic Services separation minima and conflict resolution obligations applicable to all civil aircraft operations in ICAO contracting state airspace establishing minimum separation standards between IFR and VFR aircraft that UTM AI conflict detection must support — adversarially suppressed conflict indicators that allow minimum separation violations create ICAO Annex 11 non-compliance dimensions relevant to eVTOL operations in controlled airspace; and the Advanced Air Mobility Coordination and Leadership Act P.L. 117-203 FAA leadership team and interagency coordination mandate establishing Congressional oversight of AAM integration safety performance requirements that apply across the UTM and U-space service supplier ecosystem supporting commercial eVTOL operations. Threshold: 65 for advanced air traffic management display AI — reflecting FAA UTM ConOps USS accuracy requirements, EASA U-space EU 2021/664–666 conformance monitoring service obligations, ICAO Annex 11 separation minima, and P.L. 117-203 AAM safety framework dimensions.

Integration: eVTOL AI image ingestion with Glyphward pre-scan

eVTOL AI image ingestion flows from Joby Aviation AI, Archer Aviation AI, Wisk Aero AI, Supernal AI, and Volocopter VoloCity AI onboard obstacle detection camera image pipelines, vertiport approach and landing zone surveillance camera display image interfaces, passenger biometric boarding facial recognition camera processing channels, and NASA UTM AI, AirMap AI, and Altitude Angel AI advanced air traffic management display image processing platforms into collision avoidance AI, landing clearance AI, boarding authorization AI, and airspace deconfliction AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary before AI-generated output is committed to collision avoidance manoeuvre commands, landing approach clearance records, passenger manifest boarding authorizations, or airspace conflict resolution advisories:

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"

# Urban air mobility & eVTOL AI — adversarial pixel injection in
# obstacle detection camera images, vertiport landing zone displays,
# passenger biometric boarding photos, and UTM airspace display images
# with FAA 14 CFR Part 21 Subpart H, FAA Part 135 §135.63,
# EASA CS-SC-VTOL-01, and P.L. 117-203 regulatory consequences.

# FAA §91.111 collision avoidance; FAA §91.181 course compliance;
# FAA Part 21 Subpart H Type Certificate safety case; EASA CS-SC-VTOL-01 AMC.
THRESHOLD_OBSTACLE_DETECT_AVOID_AI        = 65

# FAA Vertiport Engineering Brief VEB-1 approach surface obstruction clearance;
# EASA Vertiport Design Specification; P.L. 117-203 §4 vertiport safety.
THRESHOLD_VERTIPORT_LANDING_ZONE_AI       = 70

# FAA 14 CFR §135.63 passenger manifest; TSA Identity-Based Boarding;
# EU AI Act Art. 5(1)(d) real-time biometric identification prohibition.
THRESHOLD_PASSENGER_BIOMETRIC_BOARDING_AI = 60

# FAA UTM ConOps USS accuracy; EASA U-space EU 2021/664-666;
# ICAO Annex 11 Air Traffic Services separation minima.
THRESHOLD_AIR_TRAFFIC_MANAGEMENT_AI       = 65


class EvtolAIContext(str, Enum):
    OBSTACLE_DETECT_AVOID_AI        = "obstacle_detect_avoid_ai"         # Joby, Archer, Wisk
    VERTIPORT_LANDING_ZONE_AI       = "vertiport_landing_zone_ai"        # Joby, Supernal, Volocopter
    PASSENGER_BIOMETRIC_BOARDING_AI = "passenger_biometric_boarding_ai"  # Joby, Archer
    AIR_TRAFFIC_MANAGEMENT_AI       = "air_traffic_management_ai"        # NASA UTM, AirMap, Altitude Angel


def threshold_for(context: EvtolAIContext) -> int:
    mapping = {
        EvtolAIContext.OBSTACLE_DETECT_AVOID_AI:        THRESHOLD_OBSTACLE_DETECT_AVOID_AI,
        EvtolAIContext.VERTIPORT_LANDING_ZONE_AI:       THRESHOLD_VERTIPORT_LANDING_ZONE_AI,
        EvtolAIContext.PASSENGER_BIOMETRIC_BOARDING_AI: THRESHOLD_PASSENGER_BIOMETRIC_BOARDING_AI,
        EvtolAIContext.AIR_TRAFFIC_MANAGEMENT_AI:       THRESHOLD_AIR_TRAFFIC_MANAGEMENT_AI,
    }
    return mapping[context]


async def scan_evtol_ai_image(
    image_path: str | Path,
    context: EvtolAIContext,
    aircraft_entity_hash: str,    # SHA-256 of eVTOL serial number or UTM participant ID
    flight_operation_ref: str,    # e.g. "JOBY-FLT-2026-04881", "AIRMAP-UTM-SES-7712"
    airspace_session_id: str,     # UTM session ID or vertiport approach sequence ID
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan an eVTOL or urban air mobility AI image for adversarial injection
    payloads before forwarding to obstacle collision avoidance, vertiport
    landing zone clearance, passenger biometric boarding authorization, or
    advanced air traffic management conflict detection AI.

    Raises AdversarialEvtolAIImageError if score meets threshold:
      - OBSTACLE_DETECT_AVOID_AI:        threshold 65; FAA §91.111; EASA CS-SC-VTOL-01
      - VERTIPORT_LANDING_ZONE_AI:       threshold 70; FAA VEB-1; P.L. 117-203 §4
      - PASSENGER_BIOMETRIC_BOARDING_AI: threshold 60; FAA §135.63; EU AI Act Art. 5
      - AIR_TRAFFIC_MANAGEMENT_AI:       threshold 65; ICAO Annex 11; EU 2021/666
    """
    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": {
                "evtol_context":        context.value,
                "aircraft_entity_hash": aircraft_entity_hash,
                "flight_operation_ref": flight_operation_ref,
                "airspace_session_id":  airspace_session_id,
                "client_scan_id":       client_scan_id,
                "image_sha256":         image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "aircraft_entity_hash": aircraft_entity_hash,
        "flight_operation_ref": flight_operation_ref,
        "airspace_session_id":  airspace_session_id,
        "evtol_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_evtol_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialEvtolAIImageError(
            f"eVTOL AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"aircraft={aircraft_entity_hash} ref={flight_operation_ref}"
        )
    return result


async def write_evtol_audit_record(record: dict) -> None:
    """Persist audit record to eVTOL AI regulatory safety documentation store (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialEvtolAIImageError(Exception):
    """Raised when an eVTOL or urban air mobility AI image exceeds the adversarial injection threshold."""
    pass

Call scan_evtol_ai_image() with EvtolAIContext.OBSTACLE_DETECT_AVOID_AI before forwarding Joby Aviation AI, Archer Aviation AI, or Wisk Aero AI onboard camera image frames to collision avoidance classification AI — with aircraft_entity_hash as the SHA-256 of the eVTOL serial number and flight_operation_ref as the flight plan identifier for FAA 14 CFR §91.111 collision avoidance compliance, FAA Part 21 Subpart H Type Certificate safety case audit trail, and EASA CS-SC-VTOL-01 AMC obstacle avoidance performance documentation. Call with EvtolAIContext.VERTIPORT_LANDING_ZONE_AI for Joby Aviation AI, Supernal AI, or Volocopter VoloCity AI approach camera and landing zone surveillance display images before landing clearance authorization AI — with airspace_session_id as the approach sequence identifier for FAA Vertiport Engineering Brief VEB-1 obstruction clearance compliance, EASA Vertiport Design Specification performance documentation, and P.L. 117-203 §4 vertiport safety standard audit trail. Call with EvtolAIContext.PASSENGER_BIOMETRIC_BOARDING_AI for Joby Aviation AI or Archer Aviation AI terminal boarding lane facial recognition images before passenger identity verification and manifest documentation AI — with flight_operation_ref as the flight reservation number for FAA 14 CFR §135.63 weight and balance manifest compliance, TSA Identity-Based Boarding programme audit trail, and EU AI Act Art. 5(1)(d) biometric processing documentation. Call with EvtolAIContext.AIR_TRAFFIC_MANAGEMENT_AI for NASA UTM AI, AirMap AI, or Altitude Angel AI traffic conflict indicator display images before airspace deconfliction conflict classification AI — with airspace_session_id as the UTM session identifier for FAA UTM ConOps USS accuracy compliance, EASA U-space EU 2021/666 conformance monitoring service obligation audit trail, and ICAO Annex 11 separation minima documentation. Get early access

Coverage matrix

Related questions

Further reading

• FigStep adversarial image injection detection — technical overview of the pixel-level adversarial perturbation methodology underlying eVTOL obstacle detection camera image injection, vertiport landing zone display bypass, and UTM traffic conflict indicator suppression attacks on urban air mobility AI.
• Vision-language model security — architectural overview of multimodal AI adversarial injection covering the VLM image encoder and cross-attention layers used by eVTOL AI obstacle classification, passenger biometric boarding, and advanced air traffic management display processing pipelines.
• Free tier — 10 scans/day, no card required — start scanning eVTOL AI image inputs at development volumes before committing to a production plan; test obstacle detection camera frame injection, vertiport landing zone bypass, and biometric boarding perturbation detection without a payment method on file.
• Prompt injection in drone and UAV delivery and inspection AI — related low-altitude aviation AI injection surface covering Amazon Prime Air, Wing Aviation, Zipline, Skydio, and DJI Enterprise AI with overlapping FAA Remote ID, obstacle detection, and airspace compliance injection dimensions that share architectural parallels with eVTOL UTM air traffic management injection.
• Prompt injection in autonomous vehicle fleet safety AI — related multimodal AI injection surface for autonomous ground vehicle detect-and-avoid AI with overlapping obstacle classification suppression and safety-critical AI decision pipeline dimensions relevant to eVTOL collision avoidance AI attack surface characterisation.