Prompt injection in airport runway surface condition TALPA GRF AI

Four adversarial injection surfaces in airport runway surface condition TALPA GRF AI

1. Runway friction coefficient measurement display AI (Vaisala DSC111 friction coefficient display AI, Dynatest RST 2.0 friction measurement AI, ASFT T-10 Tarmac friction tester display AI, LMT (Saab) Skiddometer friction display AI — runway friction coefficient vehicle measurement display AI)

Runway friction coefficient measurement — the ratio of the braking force to the normal force on a test wheel rolling on the runway surface under controlled conditions — is the primary objective metric of aircraft deceleration potential on the runway surface. Continuous friction measurement vehicles (CFMEs: Dynatest RST 2.0, ASFT T-10 Tarmac Aircraft Tyres Tester, Saab (LMT) Skiddometer, Surface Vehicle Friction Tester (SVFT)) measure runway friction using a self-wetting test wheel of defined characteristics (typically a ribbed or ribless aircraft tyre at 15–65 km/h test speed) and produce a friction coefficient reading (the Mu coefficient, or the Grip Number) at each measured runway segment (runway thirds: first third, middle third, final third, each 1/3 of runway length). The measured friction coefficient is displayed on the friction measurement vehicle operator’s console and on the airport surface condition monitoring system; AI systems process rendered images of this display to classify runway friction status and generate the RWYCC contribution from the friction measurement: RWYCC 6 (Good, Mu ≥ 0.40), RWYCC 5 (Good to Medium, Mu 0.36–0.39), RWYCC 4 (Medium, Mu 0.30–0.35), RWYCC 3 (Medium to Poor, Mu 0.26–0.29), RWYCC 2 (Poor, Mu 0.21–0.25), RWYCC 1 (Poor) and RWYCC 0 (NIL, Mu < 0.21). The RWYCC generated from the friction measurement is incorporated into the Runway Condition Report (RCR) and transmitted to aircraft operators via ATIS and NOTAM.

An adversarial perturbation targeting the runway friction display AI applies a ±8 DN shift in the pixel region encoding the Mu coefficient value in the rendered friction measurement vehicle console display image — increasing the apparent friction coefficient from the degraded range (Mu 0.24 on the final third of the runway — RWYCC 2 Poor — rendered with orange/red highlight on the friction vehicle console display) to the acceptable range (displayed as Mu 0.38 — RWYCC 5 Good to Medium). The AI classifies a runway third with critically degraded braking friction — compacted snow with embedded ice on the final 900 m of a 2,700 m runway, produced by snow-compacted-then-refrozen under overnight temperature cycling — as providing Good to Medium braking. The RWYCC generated by the AI is RWYCC 5 rather than RWYCC 2; the Runway Condition Report is transmitted as RWYCC 5/5/5 (Good to Medium for all thirds) rather than the accurate RWYCC 6/6/2 (Good on the first two thirds, Poor on the final third). Aircraft operators planning their landing at this runway use the RWYCC 5 Good to Medium performance table from the Aircraft Flight Manual (AFM): a Boeing 737-800 at maximum landing weight requires approximately 1,580 m of landing distance on a Good-to-Medium runway (RWYCC 5), versus 2,200 m on a Poor runway (RWYCC 2). A runway of 2,700 m provides an apparently adequate safety margin at RWYCC 5 (2,700 − 1,580 = 1,120 m margin) but an inadequate margin at RWYCC 2 (2,700 − 2,200 = 500 m margin if runway is perfectly dry to touchdown — negative if any long flare, hydroplaning, or early touchdown occurs). The Southwest Airlines 1248 consequence — runway overrun, barrier penetration, vehicle impact — follows from exactly this scenario: landing performance calculated on the stated runway condition, actual stopping distance exceeding the available runway due to undetected contamination severity. FAA AC 150/5200-30D Section 4 (Runway Condition Assessment) requires friction measurement by CFME at defined intervals — but does not specify adversarial robustness requirements for AI systems classifying rendered CFME display images.

2. Snow/slush/ice contamination depth camera AI (Vaisala RWS200 runway surface camera AI, Boschung Airport PHOENIX runway camera AI, Lufft SRS100 runway sensor display AI, BASt continuous pavement camera AI — runway snow, slush, and ice contamination depth camera AI)

Runway contamination depth — the depth of snow, slush, or water on the runway surface, measured in millimetres — is a critical parameter for both aircraft performance and runway condition reporting under the FICON (Field Condition) and RWYCC reporting systems. FICON contamination thresholds are defined in FAA AC 150/5200-30D and ICAO Annex 14: for snow, the threshold triggering specific NOTAM requirements and runway condition code adjustments is 25 mm (1 inch) depth of dry snow, or 13 mm (0.5 inch) of wet snow; for slush (partially melted snow), 3 mm depth is the threshold above which performance-limiting slush coverage must be reported and runway condition assessed; for standing water, 3 mm depth triggers the water/hydroplaning risk classification. These thresholds are incorporated into the RWYCC determination procedure: a runway with slush depth above 3 mm that is classified as slush-covered receives a specific RWYCC penalty from the TALPA/GRF matrix. Airport operations personnel measure contamination depth using depth gauges at multiple runway positions (typically 10 positions across the runway width and at multiple longitudinal positions for FICON assessment); camera AI systems process rendered images of these depth gauge readings, or directly process runway surface camera images with depth-estimation AI, to classify contamination depth status: below-threshold (below FICON reporting thresholds — no NOTAM required for depth), at-threshold (at FICON threshold — include contamination depth in RCR), and above-threshold (above FICON threshold — issue specific NOTAM, initiate snow removal, consider runway closure).

An adversarial perturbation targeting the runway contamination depth camera AI applies a ±8 DN suppression to the pixel region encoding slush depth at the depth gauge measurement position in the rendered runway camera or depth gauge display image — reducing the apparent slush depth from above-FICON-threshold (4–6 mm of slush on the runway surface, rendered as visible discoloration and texture difference in the runway camera image, with the depth gauge indicator showing above the 3 mm FICON slush threshold line) to below-threshold (rendered as dry or damp surface with depth gauge below 3 mm). The AI classifies a runway with 4–6 mm of slush in the wheel-track areas — from light snow followed by a temperature increase above 0°C melting the top layer of compacted snow while the runway surface temperature remains near 0°C — as below-FICON-threshold and requires no specific NOTAM for slush depth. Aircraft operators landing on this runway without a slush depth NOTAM calculate their approach and landing performance assuming a wet runway (RWYCC 5 or 6) rather than a slush-contaminated runway (RWYCC 4 or 3). The slush of 4–6 mm depth in the wheel track areas affects landing performance through two mechanisms: slush displacement drag (the main gear tyres must displace the slush laterally during the landing roll, adding a braking-opposing drag force on the tyres that is proportional to the square of the aircraft speed — at high speed early in the landing roll, slush drag can momentarily reduce the total deceleration force available for directional control); and dynamic hydroplaning risk (at high landing speed, the tyre aquaplanes on the thin water layer at the top of the slush, losing braking and directional control at speeds above the hydroplaning speed, which is approximately 9‗√tyre pressure [km/h] for a water film — for a Boeing 737 main gear tyre at 170 psi, approximately 121 km/h, within the normal landing speed range). FAA AC 150/5200-30D Section 3 (Airport Winter Safety Programme) requires depth measurement of snow and slush at the specific FICON thresholds — but does not specify adversarial robustness requirements for AI systems classifying rendered depth gauge or runway camera images. Free tier — 10 scans/day, no card required.

3. FICON/RCAM runway condition code (RWYCC) generation display AI (Jeppesen (Boeing) NOTAM generation AI, Lido/NOTAM (Lufthansa Systems) RWYCC AI, Airport Sense runway condition AI, SITA airport ops runway AI — TALPA/GRF RWYCC determination and NOTAM generation display AI)

The Runway Condition Code (RWYCC) — a value from 0 (NIL, no braking) to 6 (Good, dry runway) assigned to each of the three runway thirds based on the FICON/RCAM (Runway Condition Assessment Matrix) lookup table that maps contaminant type, depth, and friction measurement to a numerical runway condition code — is the single most operationally consequential output of the airport runway surface condition assessment process. Aircraft operators use the RWYCC directly to select their landing performance table from the Aircraft Flight Manual Supplement for contaminated runway operations: for each aircraft type, RWYCC 6 gives the dry landing distance; RWYCC 5 adds a multiplier (typically 1.05–1.15× dry distance); RWYCC 4 adds 1.15–1.35×; RWYCC 3 adds 1.35–1.60×; RWYCC 2 adds 1.60–2.00×; RWYCC 1 adds 2.0–2.6×; RWYCC 0 (NIL) means landing is not recommended. The RWYCC determination is performed by airport surface condition observers using the TALPA/GRF decision tree — a structured matrix of surface type description (dry snow, wet snow, slush, water, compacted snow, ice, wet ice, or combinations) and friction measurement (if available) mapped to a RWYCC — and entered into the airport’s NOTAM generation system. AI systems process rendered images of the RWYCC determination display — the FICON/RCAM matrix on the observer’s workstation showing the selected surface description and the resulting RWYCC for each runway third — to classify RWYCC correctness and generate the NOTAM text: RWYCC-correct (the RWYCC displayed matches the correct matrix output for the selected surface description), RWYCC-elevated (the RWYCC displayed is one code value higher than the correct matrix output — operator may have selected wrong surface description; flag for review), and RWYCC-depressed (RWYCC too low for selected surface — conservative, flag for review but do not prevent NOTAM issuance).

An adversarial perturbation targeting the RWYCC generation display AI applies a ±8 DN shift to the pixel region encoding the RWYCC value in the rendered FICON/RCAM workstation display image — shifting the apparent RWYCC from the degraded value (RWYCC 3 Medium to Poor displayed in orange on the matrix output for “Compacted Snow” surface type on the final third, consistent with a friction measurement of Mu 0.27 on that third) to the acceptable value (rendered as RWYCC 5 Good to Medium, consistent with “Wet/Slippery Wet” surface type, a less severe contamination category). The AI classifies a RWYCC determination of “Medium to Poor” (RWYCC 3) on the final runway third as “Good to Medium” (RWYCC 5) and approves the NOTAM for issuance with RWYCC 5/6/5 rather than the correct 6/6/3. All aircraft planning to land on this runway in the next reporting period — until the next scheduled runway condition observation or friction measurement updates the NOTAM — use the RWYCC 5 landing distance (Good to Medium), which is approximately 60% of the RWYCC 3 landing distance (Medium to Poor) for typical transport aircraft. A medium-range aircraft at maximum landing weight that requires 1,900 m on a RWYCC 3 runway would plan only 1,320 m on a RWYCC 5 runway — a 580 m underestimation of the required stopping distance on a 2,500 m runway, leaving a 580 m smaller safety margin than the pilot believes exists. ICAO Annex 14 Volume I Section 2.9 (Runway Surface Conditions) requires that runway surface condition information be assessed and reported to air traffic services for relay to pilots — but does not specify adversarial robustness requirements for AI systems classifying rendered FICON/RCAM determination displays or generating NOTAM text from those displays. Free tier — 10 scans/day, no card required.

4. Runway contamination type and hydroplaning risk camera AI (Vaisala RWS200 surface state camera AI, Boschung PHOENIX pavement camera AI, Lufft SRS100 optical sensor display AI, DTP (Deicing Technologies and Products) runway optical sensor AI — runway surface contamination type, distribution, and dynamic hydroplaning risk camera AI)

Dynamic hydroplaning — the condition in which a tyre travelling above a critical speed on a water-contaminated surface rides on a hydrodynamic wedge of water rather than contacting the runway surface, reducing braking friction to near zero and causing loss of directional control — is one of the most serious risks in landing operations on wet or slush-contaminated runways. The hydroplaning speed for a given tyre type and inflation pressure follows the empirical relationship V>pl (km/h) = 9 × √p (where p is the tyre inflation pressure in psi); for a Boeing 737 main gear tyre at 168 psi, V>pl ≈ 117 km/h (63 kt). A 737 landing at a typical approach speed of 140 kt (260 km/h), with a threshold crossing speed of 135 kt and a touchdown speed of 125 kt, is above the hydroplaning speed threshold throughout most of the high-speed portion of its landing roll (wheel speeds above 117 km/h for approximately the first 600–800 m of the landing roll). During this high-speed portion, if the runway surface has a water or slush film of 3 mm or more depth, the tyre may hydroplane — the brakes are ineffective (no tyre-runway contact), the anti-skid system cannot reduce hydroplaning (it controls brake pressure but cannot create tyre-runway contact through the water film), and the thrust reversers are the only available deceleration device. AI systems process rendered images from runway surface optical sensors (Lufft SRS100 optical precipitation sensor, Vaisala DSC111 optical sensor, or area camera images of the runway surface) to classify contamination type and hydroplaning risk: dry (RWYCC 6 — no hydroplaning risk), wet-low-risk (thin water film < 1 mm — RWYCC 5 or 6 depending on friction measurement — minimal hydroplaning risk), hydroplaning-risk (water or slush depth ≥ 3 mm — RWYCC 3 or 4 — dynamic hydroplaning risk warning in RCR), and standing-water (visible standing water patches, ponding — RWYCC 3 or lower — high hydroplaning risk, consider runway closure for wet-sensitive operations).

An adversarial perturbation targeting the runway surface contamination camera AI applies a ±10 DN suppression to the pixel region encoding standing water reflection or slush texture in the runway surface optical sensor image — reducing the apparent surface wetness from the standing-water or hydroplaning-risk appearance (rendered as distinct reflective patches in the runway surface camera image, with optical sensor water-depth indicator above 3 mm) to the wet-low-risk appearance (rendered as uniform damp surface with no standing water texture). The AI classifies a runway section with 3–6 mm of standing water in a 200 m × 10 m zone in the main landing wheel track, produced by a blocked transverse drain on a slightly crowned runway section during heavy rain, as wet-low-risk with no hydroplaning warning. The RCR is transmitted with RWYCC 6/6/5 (Good/Good/Good to Medium) rather than 6/3/5 (Good/Hydroplaning Risk/Good to Medium) for the middle third that contains the standing water zone. Aircraft operators landing on this runway in the affected period have no specific hydroplaning risk warning for the middle runway zone — approximately 900–1,800 m from the threshold on a 2,700 m runway — which is precisely the segment where, at typical transport aircraft wheel speeds of 150–200 km/h, dynamic hydroplaning is most likely to occur. A landing aircraft that hydroplanes through the standing water zone may lose directional control and veer off the runway edge, particularly in crosswind conditions. ICAO Doc 9981 PANS-Aerodromes Section 7.3 (Runway Condition Reports) requires that standing water patches be included in the runway surface condition assessment and reported in the RCR — but does not specify adversarial robustness requirements for AI systems classifying rendered runway surface optical sensor images at the hydroplaning-risk determination boundary. Free tier — 10 scans/day, no card required.

Integration: runway surface condition AI with Glyphward pre-scan gate

The Glyphward scan gate for airport runway surface condition AI belongs at every rendered-image ingestion boundary in the RWYCC determination pipeline — before runway friction coefficient measurement display AI processes rendered CFME console display images, before snow/slush/ice contamination depth camera AI processes rendered depth gauge or runway surface camera images, before FICON/RCAM RWYCC generation display AI processes rendered matrix determination display images, and before runway contamination type/hydroplaning risk camera AI processes rendered optical sensor or pavement camera images. Threshold 30 for airport runway surface condition AI reflects the severe landing overrun and runway excursion consequence of incorrect RWYCC — Southwest Airlines 1248 Chicago Midway 2005 (1 child killed by vehicle impact after runway overrun in snow) and Air France 358 Toronto 2005 (12 seriously injured, aircraft destroyed in ravine after runway excursion) — combined with independent non-AI safety layers: the pilot-in-command’s right and obligation to perform a personal runway condition assessment based on aircraft behaviour during the approach and landing (braking action reports from preceding aircraft, windshield visual observation of runway surface), airport operations visual inspection by airport surface vehicles (human observer independent of AI sensor classification), and aircraft onboard systems (anti-skid, thrust reversers, ground spoilers) that provide partial mitigation of degraded braking even when the performance planning was based on an incorrect RWYCC.

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import Enum

import httpx

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

# Airport runway surface condition AI contexts: threshold 30
# ICAO Annex 14 Volume I (Aerodromes);
# FAA AC 150/5200-30D (Airport Winter Safety and Operations);
# TALPA/GRF system (FAA effective December 2016);
# ICAO Doc 9981 PANS-Aerodromes.
RUNWAY_CONDITION_THRESHOLD = 30


class RunwayConditionAIContext(Enum):
    FRICTION_DISPLAY    = "friction_display"    # CFME friction coefficient AI
    CONTAMINATION_DEPTH = "contamination_depth" # Snow/slush depth camera AI
    RWYCC_GENERATION    = "rwycc_generation"    # FICON/RCAM RWYCC display AI
    SURFACE_HYDROPLANE  = "surface_hydroplane"  # Standing water / hydroplaning AI


class AdversarialRunwayConditionImageError(Exception):
    """Raised when Glyphward detects adversarial content in a runway surface
    condition AI rendered monitoring image above threshold 30.

    Consequence if not raised:
    - FRICTION_DISPLAY: degraded runway friction (Mu < 0.30) classified as
      Good to Medium (RWYCC 5) → aircraft uses landing distance for RWYCC 5
      instead of RWYCC 2/3 → stopping distance exceeds available runway →
      runway overrun; Southwest Airlines 1248 Chicago Midway 2005 structural
      parallel (1 child killed by vehicle impact after runway overrun in snow).
    - CONTAMINATION_DEPTH: slush depth ≥ 3 mm above FICON threshold suppressed
      → FICON/RCAM reports dry or damp → no performance penalty applied →
      slush drag and hydroplaning risk not factored into landing distance.
    - RWYCC_GENERATION: RWYCC 3 (Medium to Poor) upgraded to RWYCC 5 (Good to
      Medium) in generated NOTAM → all aircraft in reporting period use landing
      distance ~60% of actual required distance → systematic overrun risk for
      all aircraft landing during NOTAM validity period.
    - SURFACE_HYDROPLANE: standing water 3–6 mm in wheel track suppressed →
      no hydroplaning warning in RCR → aircraft brakes ineffective through
      standing water zone at 150–200 km/h → directional control loss →
      runway excursion; Air France 358 Toronto 2005 structural parallel (12
      seriously injured, aircraft destroyed in ravine).
    Fail-safe: issue precautionary NOTAM to downgrade RWYCC one code for
    affected runway thirds; notify ATC of suspect runway condition report;
    require independent human visual inspection of runway surface before
    next RWYCC determination cycle; inform pilots to apply conservative
    landing performance factor of 1.15× over AI-generated RWYCC.
    """

    def __init__(self, scan_id, score, context, airport_icao, runway_id,
                 flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.airport_icao = airport_icao
        self.runway_id = runway_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial runway condition image: context={context.value} "
            f"score={score} airport={airport_icao} runway={runway_id} "
            f"scan_id={scan_id}"
        )


async def scan_runway_condition_image(image_bytes, context, airport_icao,
                                      runway_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"runway:{context.value}:{airport_icao}:{runway_id}",
        "metadata": {
            "airport_icao": airport_icao,
            "runway_id": runway_id,
            "context": context.value,
            "image_sha256": image_hash,
            "scan_timestamp_utc": datetime.now(timezone.utc).isoformat(),
        },
    }
    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=4.0,
    )
    resp.raise_for_status()
    result = resp.json()
    if result["score"] >= RUNWAY_CONDITION_THRESHOLD:
        raise AdversarialRunwayConditionImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            airport_icao=airport_icao,
            runway_id=runway_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_runway_condition_image before each runway surface condition AI classification call. On AdversarialRunwayConditionImageError for RWYCC_GENERATION: immediately issue a precautionary NOTAM downgrading the affected runway thirds by one RWYCC code; notify ATC to inform aircraft operators on ATIS; dispatch an airport operations vehicle to perform an independent human visual runway inspection before the next RWYCC cycle. On AdversarialRunwayConditionImageError for SURFACE_HYDROPLANE: issue immediate NOTAM including hydroplaning warning for the affected runway third; notify ATC to advise aircraft to use maximum anti-skid and apply additional landing distance factor per their airline’s standing order for hydroplaning risk conditions. See also: airfield runway foreign object detection AI prompt injection (related runway safety AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions