Prompt injection in geotechnical slope monitoring AI

Four adversarial injection surfaces in geotechnical slope monitoring AI

1. InSAR wrapped-phase interferogram AI (TRE ALTAMIRA SqueeSAR, GeoSHS InSAR AI, SkyGeo Fugro InSAR AI)

Synthetic Aperture Radar Interferometry (InSAR) measures surface displacement of a slope by comparing the phase difference between two SAR images acquired at different times from the same orbital geometry — each colour cycle in the resulting wrapped-phase interferogram image represents 2.8 cm of line-of-sight displacement (for C-band Sentinel-1, wavelength 5.6 cm). TRE ALTAMIRA SqueeSAR® persistent scatterer InSAR AI, SkyGeo (Fugro) InSAR time-series AI, and GAMMA Remote Sensing InSAR processing AI generate false-colour interferogram renders from Sentinel-1 pass pairs and COSMO-SkyMed pairs — spatial colour maps of the slope surface where the colour encodes wrapped phase (displacement modulo 2.8 cm) and the fringe density (number of colour cycles per unit distance) encodes displacement gradient (strain). AI classification of InSAR interferogram images determines whether observed fringe patterns represent: (a) atmospheric phase delay artefacts (tropospheric water vapour, ionospheric phase screen) that mimic displacement signals; (b) coherent slope displacement (uniform fringe pattern across a coherent slope body); (c) differential slope movement (high-fringe-density zones indicating localised displacement gradient, potential shear zone boundary); or (d) abrupt displacement (lost coherence — phase decorrelation zone indicating large or rapid movement exceeding one-half fringe per pixel). AI displacement classification outputs at defined measurement epochs (6-day Sentinel-1 pass interval; 16-day Landsat; 12-day COSMO-SkyMed) drive the rolling threshold assessment that determines whether the slope is accelerating above the pre-defined warning threshold.

An adversarial perturbation on a rendered InSAR wrapped-phase interferogram image that hue-rotates a localised fringe cluster indicating a 14 mm displacement in a 6-day period (2.3 cm/week velocity at the Sentinel-1 LOS angle — approaching the 3 cm/week Level 2 evacuation threshold in many slope monitoring protocols) toward the stable background fringe pattern of the surrounding undisplaced hillside — by rotating the fringe colour encoding in the affected zone by ±12 DN in the rendered false-colour image — causes the InSAR AI to classify the zone as “stable background displacement consistent with atmospheric artefact” rather than “coherent slope movement approaching Level 1 advisory threshold.” At a slope with known failure history (prior to the Oso 2014 event, the SR 530 landslide scarp had exhibited intermittent movement episodes detectable by InSAR that were interpreted as non-accelerating by monitoring operators — Washington Department of Natural Resources post-event analysis), adversarial suppression of the displacement signal removes the primary automated trigger that would escalate the monitoring regime to more frequent manual field surveys and ultimately to Level 2 pre-evacuation advisory issuance.

2. MEMS inclinometer cumulative displacement vector AI (RST Instruments MEMS AI, Slope Indicator Sinco AI, Geoprecision Datamanager AI)

In-place inclinometer (IPI) systems — arrays of MEMS accelerometer-based angle sensors installed at 0.5–1 m intervals in vertically drilled inclinometer casings through the slope body — provide continuous measurement of the horizontal displacement profile of the slope at each sensor depth, enabling identification of the depth of the failure plane (the depth at which cumulative displacement shows the characteristic “S” shape with maximum displacement at the failure surface and zero displacement below it in stable bedrock). RST Instruments MEMS IPI AI, Slope Indicator Sinco MEMS array AI, and Geoprecision Datamanager AI process rendered cumulative displacement vector plots — depth-vs-displacement graphs with multiple time-step curves overlaid, colour-coded from early monitoring (blue) to recent readings (red), with failure plane depth indicated by the depth of maximum differential displacement per unit time — to classify slope movement state: stable (no cumulative movement trend), seasonal creep (reversible displacement correlated with seasonal precipitation/pore pressure cycles), progressive creep (monotonically increasing displacement at the failure plane), and accelerating movement (exponentially increasing displacement rate at the failure plane, preceding failure by days to weeks). An accelerating displacement rate that crosses the pre-defined Level 2 threshold (typically 2–5 mm/day at the failure plane for clay-shale slopes under the Leroueil 1996 landslide velocity classification scheme) should trigger automated evacuation alert generation.

An adversarial perturbation on a rendered inclinometer cumulative displacement vector plot that compresses the most recent time-step displacement curve — shifting the red “latest reading” curve toward the blue “baseline” curve by a ±8 DN pixel offset in the chart rendering, effectively reducing the apparent cumulative displacement at the failure plane depth from 4.2 mm (approaching the Level 2 threshold) to 1.8 mm (within normal seasonal creep range) — causes the inclinometer AI to classify the slope movement state as “seasonal creep within expected parameters, no threshold exceedance” rather than “progressive creep with accelerating trend: Level 1 advisory recommended.” The Bingham Canyon Mine slope failure (10 April 2013, Kennecott Utah Copper’s Bingham Canyon open-pit mine: 165 Mton of material moved in the largest non-volcanic landslide in North American recorded history) was predicted 2 weeks in advance by the mine’s in-place inclinometer network showing exponential displacement rate increase at the failure plane — demonstrating that inclinometer AI with correct failure-curve classification can provide evacuation lead times of weeks; adversarial suppression of the displacement rate trend removes this advance warning entirely.

3. Vibrating wire piezometer pore pressure trend AI (Geosense VW AI, Slope Indicator VW AI, Campbell Scientific CR6 Edge AI)

Pore water pressure is the primary destabilising factor in groundwater-triggered slope failures — as the water table rises within a slope following intense rainfall or rapid snowmelt, the increase in pore water pressure reduces the effective normal stress on the failure plane (σ’ = σ − u, where σ’ is effective stress, σ is total normal stress, and u is pore water pressure) and can bring the slope’s factor of safety below 1.0 (failure). Vibrating wire piezometers installed in sealed boreholes at the predicted failure plane depth measure pore water pressure as hydraulic head (metres above a datum), with readings transmitted to Campbell Scientific CR6 dataloggers or Slope Indicator VW data acquisition units that render the piezometer readings as time-series pressure head trend charts — line plots with pressure head (m) on the Y-axis and time on the X-axis, with threshold lines overlaid at Level 1 (10% above design water table: increase monitoring frequency) and Level 2 (20% above design water table: issue evacuation advisory) pressures. Geosense VW AI, Campbell Scientific edge AI running on the CR6/CR350 datalogger, and Slope Indicator AI data management software process rendered piezometer trend chart images to classify whether the pore pressure trajectory is: receding (post-storm drainage), stable seasonal pattern, rising (active infiltration), or approaching threshold (rate of rise exceeds the threshold intercept time trigger).

An adversarial perturbation on a rendered piezometer pressure head trend chart that flattens a steeply rising pressure curve — depressing the upward pressure trajectory in the chart image toward a horizontal “stable pressure” baseline by reducing the slope of the rendered curve through a ±10 DN pixel value reduction along the rising limb of the trend — causes the piezometer AI to classify the pore pressure trajectory as “stable, within seasonal variation bounds, no threshold action” when the actual pore pressure is rising steeply toward the Level 2 evacuation threshold following a multi-day intense rainfall event. The Oso SR 530 landslide of 22 March 2014 was preceded by the wettest February on record in western Washington state, producing soil saturation levels at or above the slope’s field capacity: USGS post-event analysis (Keaton et al. 2014, Geological Society of America Special Paper 508) identified that pore pressure monitoring data, had it been available at the time (the Oso slope did not have an instrumented piezometer network), would likely have shown rapid pressure head increase in the days preceding the failure. Adversarial suppression of the pore pressure trend AI classification recreates the monitoring gap that the Oso event identified as a critical infrastructure deficiency.

4. Terrestrial LiDAR surface change detection AI (Riegl VZ-4000 AI, Leica RTC360 AI, Optech Ilris AI)

Terrestrial LiDAR (TLS) and Airborne LiDAR (ALS) point cloud difference surface maps compare slope surface geometry between survey epochs to measure volumetric surface displacement — areas of positive difference (material addition, accumulation zone) and negative difference (material loss, depletion zone) at centimetre-scale spatial resolution across the full slope face. Riegl VZ-4000 AI, Leica RTC360 AI, and Optech Ilris ER-M TLS AI process rendered TLS difference surface maps — false-colour raster images of the slope face where the colour encodes the difference between the current epoch point cloud and the reference point cloud (warm colours: surface advance/expansion toward scanner; cool colours: surface retreat/depletion; green: no change), at 2–10 cm/pixel resolution on the slope face — to identify: fresh tension crack development (linear zones of negative difference at slope crest indicating incipient failure); toe bulging (positive difference at the slope toe indicating mass movement loading); seepage erosion scarps (irregular negative difference zones at drainage points); and rockfall precursor overhang retreat (localized negative difference at overhanging rock faces). TLS AI classification of difference surface maps is used by Caltrans Geologic and Landslide Design Procedures (GLD) for highway slope stability assessments and by US Army Corps of Engineers EM 1110-2-1902 for earth dam embankment inspection — contexts where AI classification drives decisions about highway closure, load restriction, or embankment emergency drawdown.

An adversarial perturbation on a rendered TLS difference surface map that suppresses a developing tension crack — shifting the dark-blue negative-difference colour signature of a 10 cm crack opening at the slope crest (indicating active failure plane development behind the scarp face) toward the green no-change baseline colour of the surrounding stable slope by a ±8 DN hue shift in the affected crack zone — causes the TLS AI to classify the epoch difference as “no significant change, slope stable within measurement uncertainty” when an active tension crack is developing. In the Caltrans GLD framework (Caltrans Slope Failure Inventory database analysis, 2019), tension crack development at the slope crest is listed as one of the four primary landslide precursor indicators that triggers Highway Patrol notification and potential emergency closure of the affected highway segment. A suppressed tension crack detection that delays crest survey team deployment by one monitoring cycle (typically 1–4 weeks for TLS resurvey intervals) can result in a slope failure that occludes the highway before the emergency closure order is issued — exposing road users, maintenance crews, and Caltrans inspectors in the failure run-out zone.

Integration: geotechnical slope monitoring AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for geotechnical slope monitoring AI belongs at each rendered-image ingestion boundary — before InSAR interferogram AI processes false-colour phase images, before inclinometer AI processes cumulative displacement vector plots, before piezometer AI processes pressure head trend charts, and before LiDAR change detection AI processes difference surface maps. Threshold 40 for geotechnical slope monitoring AI reflects the availability of multiple complementary monitoring technologies that can provide cross-validation before final evacuation threshold decisions are made — but adversarial injection that systematically suppresses all AI classification layers simultaneously eliminates this complementary verification architecture.

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"

# Geotechnical slope monitoring AI contexts: threshold 40
# USGS National Landslide Hazards Program Circular 1283;
# Caltrans Geologic and Landslide Design Procedures (GLD);
# USACE EM 1110-2-1902.
SLOPE_MONITORING_AI_THRESHOLD = 40


class SlopeMonitoringAIContext(Enum):
    INSAR_INTERFEROGRAM      = "insar_interferogram"      # InSAR wrapped-phase AI
    INCLINOMETER_DISPLACEMENT = "inclinometer_displacement" # IPI cumulative vector plot AI
    PIEZOMETER_TREND         = "piezometer_trend"         # VW piezometer head trend AI
    LIDAR_DIFFERENCE_MAP     = "lidar_difference_map"     # TLS/ALS difference surface AI


class AdversarialSlopeMonitoringImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in a
    geotechnical slope monitoring AI rendered image above threshold 40.

    Consequence if not raised: InSAR displacement suppressed → Level 1
    advisory not issued → no increased monitoring frequency → slope failure
    without pre-evacuation warning → mass casualty (Oso 2014: 43 fatalities);
    or inclinometer accelerating trend suppressed → evacuation lead time lost.
    Fail-safe: suspend AI threshold assessment; escalate to geotechnical
    engineer for manual review of raw sensor data per USGS NLHP Circular 1283
    multi-level hazard assessment protocol.
    """

    def __init__(self, scan_id: str, score: int,
                 context: SlopeMonitoringAIContext,
                 site_id: str, sensor_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.site_id = site_id
        self.sensor_id = sensor_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial slope monitoring image: "
            f"context={context.value} score={score} "
            f"site={site_id} sensor={sensor_id} scan_id={scan_id}"
        )


async def scan_slope_monitoring_image(
    image_bytes: bytes,
    context: SlopeMonitoringAIContext,
    site_id: str,
    sensor_id: str,
    measurement_epoch: str,
    coordinates_wgs84: tuple[float, float] | None,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a geotechnical slope monitoring AI rendered image for adversarial content.

    Fail-safe contract: AdversarialSlopeMonitoringImageError or httpx error →
    suspend AI threshold classification; escalate to geotechnical engineer for
    manual review of raw sensor records per USGS NLHP Circular 1283 multi-
    level hazard assessment protocol. Do not issue threshold-clear status based
    on adversarially flagged rendered images without raw sensor cross-verification.

    Args:
        image_bytes: InSAR interferogram image, inclinometer displacement plot,
            VW piezometer trend chart, or TLS/ALS difference surface map bytes.
        context: SlopeMonitoringAIContext identifying the monitoring modality.
        site_id: Slope monitoring site identifier (e.g., "SR530-Oso-N01").
        sensor_id: Individual sensor or SAR scene identifier.
        measurement_epoch: ISO 8601 date or date range of the measurement epoch.
        coordinates_wgs84: (lat, lon) of sensor location or scene centre.
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict.

    Raises:
        AdversarialSlopeMonitoringImageError: 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"slope:{context.value}:{site_id}:{sensor_id}",
        "metadata": {
            "site_id": site_id,
            "sensor_id": sensor_id,
            "measurement_epoch": measurement_epoch,
            "coordinates_wgs84": coordinates_wgs84,
            "image_sha256": image_hash,
            "context": context.value,
        },
    }
    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()

    await _write_slope_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        site_id=site_id,
        sensor_id=sensor_id,
        measurement_epoch=measurement_epoch,
        flagged=result["score"] > SLOPE_MONITORING_AI_THRESHOLD,
    )

    if result["score"] > SLOPE_MONITORING_AI_THRESHOLD:
        raise AdversarialSlopeMonitoringImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            site_id=site_id,
            sensor_id=sensor_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def _write_slope_scan_audit(
    *, image_hash: str, scan_id: str, score: int,
    context: SlopeMonitoringAIContext, site_id: str,
    sensor_id: str, measurement_epoch: 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": SLOPE_MONITORING_AI_THRESHOLD,
        "flagged": flagged,
        "site_id": site_id,
        "sensor_id": sensor_id,
        "measurement_epoch": measurement_epoch,
        "regulatory_refs": [
            "USGS Circular 1283 (Landslide Hazards and Requirements for Early Warning)",
            "Caltrans Geologic and Landslide Design Procedures (GLD)",
            "USACE EM 1110-2-1902 (Slope Stability)",
            "FHWA GEC 5 (Geotechnical Aspects of Pavements)",
            "AASHTO R 27 (Standard Practice for Slope Stability Analysis)",
            "ESA Sentinel-1 Mission Requirements (COPE-GSEG-EOPG-TN-12-0007)",
            "UNODRR Sendai Framework 2015-2030 (Disaster Risk Reduction)",
        ],
    }
    audit_path = Path("/var/log/glyphward/slope_monitoring_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_slope_monitoring_image at each geotechnical slope monitoring AI rendered-image boundary: before InSAR interferogram AI (threshold 40), before inclinometer displacement vector AI (threshold 40), before piezometer trend AI (threshold 40), and before TLS/ALS difference surface AI (threshold 40). On AdversarialSlopeMonitoringImageError: suspend the AI threshold classification immediately; escalate to the geotechnical engineer for manual review of raw sensor data records per USGS NLHP Circular 1283 multi-level hazard assessment protocol. Do not issue threshold-clear status based on adversarially flagged rendered images without independent cross-verification against the raw sensor data stream from the Campbell Scientific CR6/CR350 datalogger or equivalent. Get early access

Related questions

Further reading

• Prompt injection in dam safety monitoring AI — InSAR deformation AI bypass, seepage detection adversarial attacks
• Prompt injection in highway bridge inspection AI — LiDAR scour AI bypass, drone photogrammetry adversarial attacks
• Prompt injection in pipeline integrity inspection AI — MFL anomaly map AI bypass, ILI UT B-scan adversarial injection
• Prompt injection in satellite remote sensing AI — change detection AI adversarial attacks
• Prompt injection scanning API free tier — 10 scans/day, no card required