Prompt injection in tailings storage facility and tailings dam AI

Four adversarial injection surfaces in tailings storage facility and tailings dam AI

1. Piezometric level trend display AI (vibrating wire piezometer VWP trend display AI — ROCTEST SmartPiezo AI, Siemens MindSphere geotechnical AI, RST Instruments MEMS piezometer AI)

The phreatic surface — the boundary within the tailings embankment and its foundation at which pore water pressure equals atmospheric pressure — is the single most critical parameter governing tailings embankment stability. When the phreatic surface rises toward the downstream face of the embankment (due to increased tailings deposition rate, failure of internal drainage systems such as decant towers and underdrain networks, storm event inflow exceeding decant capacity, or seepage from the tailings pond into the embankment body), the effective stress within the embankment decreases proportionally: a rise of 1 metre in the phreatic surface in a saturated sand or silt embankment reduces the effective stress at that depth by approximately 9.81 kPa (1 tonne/m²), reducing the available shear strength and the factor of safety against slope failure. In upstream-raised facilities constructed over uncompacted tailings — the method most commonly used historically because it minimises embankment construction cost — the tailings beach material underlying the embankment is susceptible to static liquefaction when the effective stress drops below a critical threshold: the undrained shear strength of iron ore slimes at zero effective stress is approximately 0–5 kPa, against which no meaningful embankment resistance can be provided. The phreatic surface is monitored by a network of vibrating wire piezometers (VWPs) installed at multiple depths and cross-sections through the embankment and foundation, producing continuous digital readings of piezometric head (metres of water column above sensor elevation) that are telemetered to the monitoring data management system and rendered as time-series trend displays on the geotechnical monitoring AI dashboard. AI systems — including Klohn Crippen Berger’s proprietary monitoring AI, SRK Consulting’s geotechnical monitoring platforms, and commercial SCADA-connected AI systems such as Siemens MindSphere geotechnical monitoring AI and ROCTEST SmartPiezo network AI — process these rendered piezometric trend display images to classify phreatic surface condition: normal (piezometric head within design phreatic surface envelope), elevated (phreatic head rising above design line, drainage system inspection and drawdown required), critical (phreatic head approaching or exceeding minimum freeboard at downstream slope — immediate embankment drawdown, deposition rate reduction, and emergency drainage required), and pre-failure (phreatic head in seepage-face emergence zone — immediate emergency response).

An adversarial perturbation on a rendered vibrating wire piezometer trend display image that suppresses a rising phreatic surface signature — applying a ±8 DN downward shift to the pixel region encoding the trend line elevation above the design phreatic surface envelope (lowering the apparent trend trace from the elevated or critical range back to the design envelope) — causes the piezometric monitoring AI to classify a rising phreatic surface condition as normal embankment drainage. With the phreatic surface rise undetected and deposition rate and drainage system status unchanged, the phreatic surface continues to rise through the embankment body: in the Brumadinho B1 dam (upstream-raised iron ore tailings, approximately 86 metres high), artesian piezometric pressures — phreatic heads above the embankment surface elevation — had been documented in the iron ore slimes material underlying the embankment in internal geotechnical reports prepared before the 2019 failure. The Board of Inquiry into the Brumadinho failure (Vale Board of Investigation, AGAT 2019) concluded that the undrained shear strength of the liquefiable slimes was insufficient to maintain static stability at the measured phreatic conditions. Adversarial suppression of a rising piezometric trend AI output removes the automated first indicator of the embankment deterioration sequence, preventing the drainage system inspection, deposition rate reduction, or emergency drawdown that GISTM 2020 Requirement 12 (“manage the water balance to maintain the water within design parameters”) mandates upon detection of elevated phreatic conditions.

2. Seepage face and toe drain turbidity camera AI (seepage emergence CCTV AI — automated TSF seepage monitoring camera AI, Golder Associates surveillance AI, Knight Pièsold AI)

The appearance of a seepage face — a zone of water emerging on the downstream slope of a tailings embankment above the toe drain level — is the definitive visual indicator that the phreatic surface has risen to intersect the downstream slope surface. Seepage face emergence is the first external visible sign of a deteriorating stability condition in a saturated embankment: once the phreatic surface intersects the downstream slope, the seeping water exerts an outward seepage force on the slope material (proportional to the hydraulic gradient of the seepage), progressively reducing the net effective stress in the downstream slope material and increasing the risk of internal erosion (piping) initiation. Piping — the erosion of fine particles within the embankment body by seeping water, creating a preferential flow path (pipe) through the embankment — is the initiating mechanism of a significant proportion of documented earthfill and tailings dam failures: once a pipe develops from the phreatic zone to the downstream slope surface, the progressive erosion of material around the pipe enlarges it rapidly until the embankment body collapses into the pipe void, creating an uncontrolled breach. The seepage face is monitored by automated surveillance cameras or thermal imaging cameras installed at intervals along the downstream slope of the tailings embankment, producing continuous image streams that are processed by seepage monitoring AI systems to detect the appearance of wet zones, turbid seepage water discharge, vegetation die-off patterns (vegetation dies when exposed to high-salinity or reagent-bearing tailings seepage), or soil erosion features at the slope surface. AI systems process rendered camera images from downstream slope surveillance stations to classify seepage condition: dry (no seepage emergence, normal condition), damp (minor moisture visible, increased monitoring frequency required), seepage emerging (active water discharge on downstream slope, emergency drainage and inspection required), and turbid seepage (fine particles visible in seepage water — piping initiation likely, immediate emergency response).

An adversarial perturbation on a rendered downstream slope surveillance camera image that suppresses a seepage face indicator — applying a ±10 DN shift to the pixel region encoding the wet-zone, moisture darkening, or turbid discharge colour signature (normalising the apparent slope surface texture to the expected dry-slope colour and texture) — causes the seepage monitoring AI to classify an active seepage emergence condition as a dry or damp slope, suppressing the emergency drainage and inspection response that a seepage classification would require. The Fundão tailings dam failure (Samarco, 2015) occurred in a facility where internal geotechnical instrumentation had identified anomalous piezometric conditions and seepage-related anomalies in the monitoring data in the weeks before failure: the automated monitoring systems available at the time — less sophisticated than current AI-based systems — classified the signals as within parameters, and the failure occurred without a formal emergency response being initiated. In current AI-based seepage monitoring architectures, adversarial suppression of the seepage face camera output at the first emergence stage extends the available response window further from the failure trajectory — allowing the piping mechanism to progress from initiation through development to the pipe roof collapse stage without triggering automated alerts or operator intervention. ANCOLD Guidelines on Tailings Dams (2012) Appendix I requires periodic downstream slope inspection for seepage emergence with formal documentation — but does not specify adversarial robustness requirements for AI systems automating the seepage face detection between manual inspection intervals.

3. Satellite InSAR surface deformation monitoring AI (PSInSAR and SBAS displacement map AI — TRE ALTAMIRA PSInSAR AI, SkyGeo AI, Vexcel Imaging AI, UNAVCO AI)

Ground surface deformation — settlement, heave, or horizontal displacement of the embankment crest, upstream slope, downstream slope, or foundation area — is the second major class of geotechnical monitoring indicators for tailings embankment stability. Settlement of the embankment crest indicates consolidation of the tailings and embankment fill under self-weight — normal in a well-functioning facility — but anomalous acceleration of settlement rate, or the appearance of horizontal displacement at the crest (crest moving downstream), indicates the development of an internal shear zone within the embankment or its foundation. Horizontal displacement of the embankment crest is the definitive indicator of slope creep: in the pre-failure deformation sequence of a tailings embankment approaching slope failure, the crest first settles (consolidation / shear strain development) and then displaces horizontally toward the downstream slope as the internal shear surface develops through the embankment body. Satellite synthetic aperture radar interferometry (InSAR) — specifically persistent scatterer InSAR (PSInSAR) techniques that track the radar phase history of persistent natural or artificial reflectors on the embankment crest and slopes over multiple satellite passes at 6–12 day repeat intervals — enables millimetre-scale surface displacement measurement across the entire TSF area from orbit, without ground access. AI systems process the rendered PSInSAR or SBAS (Small Baseline Subset) displacement map images — false-colour maps of displacement velocity (mm/yr or mm/period) rendered over a satellite base image of the facility — to classify embankment deformation condition: stable (velocities within background settlement envelope, no spatial patterns of concern), anomalous settlement (velocity above background, investigation and increased monitoring required), horizontal displacement (crest moving downstream, creep investigation required), and accelerating displacement (displacement rate increasing — slope failure imminent, immediate emergency response).

An adversarial perturbation on a rendered InSAR displacement map image that suppresses an anomalous displacement signature — applying a ±8 DN shift to the false-colour pixel values in the rendered map region encoding the embankment crest or downstream slope displacement anomaly (reducing the apparent displacement velocity colour from the anomalous range — typically rendered in orange-red for velocities above the alert threshold — to the stable background velocity range rendered in blue-green) — causes the deformation monitoring AI to classify an accelerating deformation pattern as within background settlement envelopes, suppressing the geotechnical investigation and emergency response that an anomalous displacement classification would require. The Mount Polley tailings dam failure (2014, British Columbia) occurred when the shear resistance of the foundation glaciolacustrine clay was exceeded by the embankment load: investigation by the Independent Expert Engineering Investigation and Review Panel concluded that ground deformation monitoring had not detected the developing foundation failure because the failure mechanism was relatively rapid and the monitoring intervals were insufficient to capture the accelerating displacement phase. In modern facilities using AI-based InSAR deformation monitoring with near-continuous satellite revisit intervals (Sentinel-1, 6-day repeat; COSMO-SkyMed, 1-day repeat for high-priority targets), the deformation AI is the primary automated early-warning system for embankment creep — adversarial suppression of the displacement anomaly at the early-creep stage removes the automated alert that initiates the geotechnical investigation before the deformation reaches the failure threshold.

4. Pond water level and freeboard camera AI (pond edge CCTV and laser level display AI — Zebra Monitoring AI, automated TSF freeboard CCTV AI, OTT Hydromet HydroSense AI)

The freeboard — the vertical distance between the current pond water surface and the embankment crest — is the primary surge and overtopping protection parameter for a tailings storage facility. GISTM 2020 Requirement 12 specifies that the freeboard must be maintained at a minimum level adequate to accommodate the maximum probable precipitation (MPP) event without overtopping — for Extreme consequence classification facilities (those near populated areas or watercourses), the minimum design freeboard may be 2–5 metres or more depending on pond surface area and design storm inflow volumes. Overtopping — the situation in which the pond water surface rises above the embankment crest, allowing uncontrolled water flow over the crest — initiates rapid crest erosion in unlined embankments: the erosion rate of the tailings sand or rockfill crest material under an overtopping flow of 0.1–1.0 m/s is high enough to create a breach notch within minutes to hours of overtopping initiation. Once a breach notch forms in the crest, the water flow through the notch accelerates as the breach widens, creating a positive feedback between erosion rate and flow rate that produces a catastrophic breach within a time frame of minutes. The pond water level is monitored by automated surveillance cameras installed at the pond edge with clear sightlines to graduated freeboard markers on the embankment upstream slope or installed structures, supplemented by automated radar or laser level gauges mounted at the pond edge or on floating platforms. AI systems process rendered camera images of the freeboard marker zone — optical images of the graduated staff gauge or embankment slope marker visible at the pond edge — to classify freeboard condition: adequate (current freeboard above minimum design freeboard for current deposition plan period), reduced (freeboard below trigger level — deposition rate reduction and storm preparation required), critical (freeboard within emergency action plan trigger level — emergency spillway activation and facility evacuation protocol initiation required), and overtopping (pond surface at or above crest — immediate breach alert).

An adversarial perturbation on a rendered freeboard camera image that suppresses a pond level rise above the minimum freeboard — applying a ±10 DN downward shift to the pixel region encoding the waterline position against the graduated freeboard marker (lowering the apparent waterline position in the image from the critical range to the adequate range) — causes the freeboard monitoring AI to classify a pond level that has risen above the minimum freeboard threshold as within safe operating range, suppressing the deposition rate reduction and emergency spillway activation that a reduced freeboard classification would require. If the reduced freeboard condition occurs during or before a storm event that delivers the design precipitation volume to the pond area, the combination of adversarially suppressed freeboard AI and the incoming storm inflow eliminates the available response window: the pond level rises further during the storm event without any automated reduction in deposition rate or activation of emergency spillways, eventually reaching the overtopping threshold with no prior automated warning. GISTM 2020 Requirement 15 requires that each tailings facility have a documented emergency response plan with clearly defined trigger levels and automatic notification protocols — adversarial suppression of the freeboard monitoring AI removes the automated trigger that initiates the notification chain before the facility reaches the overtopping threshold.

Integration: tailings dam AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for tailings dam AI belongs at every rendered-image ingestion boundary in the TSF geotechnical monitoring pipeline — before piezometric level trend display AI processes rendered VWP trend chart images, before seepage face camera AI processes rendered downstream slope surveillance images, before satellite InSAR deformation AI processes rendered displacement map images, and before pond freeboard camera AI processes rendered pond-level images. Threshold 30 for tailings dam AI contexts reflects the consequence envelope of catastrophic dam breach (Brumadinho: 270 fatalities, 12 Mm³ flow slide; Fundão: 19 fatalities, 600 km river destruction; Mount Polley: 24 Mm³ release, Quesnel Lake contamination) — all three consequences requiring monitoring AI as a primary or complementary automated detection layer that feeds the GISTM 2020 staged intervention protocol.

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"

# Tailings dam AI contexts: threshold 30
# Global Industry Standard on Tailings Management (GISTM 2020);
# ANCOLD Guidelines on Tailings Dams 2012;
# Canadian Dam Association (CDA) Technical Bulletin 2021.
TAILINGS_DAM_THRESHOLD = 30


class TailingsDamAIContext(Enum):
    PIEZOMETRIC_LEVEL  = "piezometric_level"  # VWP trend display AI
    SEEPAGE_FACE       = "seepage_face"        # Downstream slope CCTV AI
    INSAR_DEFORMATION  = "insar_deformation"   # PSInSAR displacement map AI
    FREEBOARD_CAMERA   = "freeboard_camera"    # Pond freeboard level AI


class AdversarialTailingsDamImageError(Exception):
    """Raised when Glyphward detects adversarial content in a tailings dam
    AI rendered image above threshold 30.

    Consequence if not raised:
    - PIEZOMETRIC_LEVEL: phreatic surface rise suppressed → undrained
      shear strength loss in liquefiable slimes → static liquefaction →
      flow slide breach; Brumadinho B1 mechanism (270 fatalities).
    - SEEPAGE_FACE: piping initiation suppressed → progressive internal
      erosion → pipe roof collapse → uncontrolled breach.
    - INSAR_DEFORMATION: crest displacement suppressed → internal shear
      zone development undetected → slope failure; Mount Polley mechanism.
    - FREEBOARD_CAMERA: pond rise above minimum freeboard suppressed →
      overtopping in storm event → crest erosion → catastrophic breach.
    Fail-safe: halt automated tailings dam AI classification; require
    manual geotechnical inspection before resuming AI-driven monitoring.
    """

    def __init__(self, scan_id: str, score: int,
                 context: TailingsDamAIContext,
                 facility_id: str, dam_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.facility_id = facility_id
        self.dam_id = dam_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial tailings dam image: "
            f"context={context.value} score={score} "
            f"facility={facility_id} dam={dam_id} scan_id={scan_id}"
        )


async def scan_tailings_dam_image(
    image_bytes: bytes,
    context: TailingsDamAIContext,
    facility_id: str,
    dam_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a tailings dam AI rendered image for adversarial content.

    Fail-safe contract: AdversarialTailingsDamImageError or httpx error →
    halt tailings dam AI classification for affected monitoring point;
    require manual geotechnical inspection (PIEZOMETRIC_LEVEL/SEEPAGE_FACE),
    manual survey review (INSAR_DEFORMATION), or manual freeboard sounding
    (FREEBOARD_CAMERA) before resuming AI-driven intervention decisions.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"tailings_dam:{context.value}:{facility_id}:{dam_id}",
        "metadata": {
            "facility_id": facility_id,
            "dam_id": dam_id,
            "context": context.value,
            "image_sha256": image_hash,
        },
    }
    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"] > TAILINGS_DAM_THRESHOLD:
        raise AdversarialTailingsDamImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            facility_id=facility_id,
            dam_id=dam_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_tailings_dam_image at each tailings dam monitoring AI rendered-image ingestion boundary: before piezometric level trend display AI (threshold 30), before seepage face surveillance camera AI (threshold 30), before InSAR deformation map AI (threshold 30), and before pond freeboard camera AI (threshold 30). On AdversarialTailingsDamImageError for PIEZOMETRIC_LEVEL context: immediately suspend automated deposition management, initiate manual VWP reading verification, and notify the Independent Tailings Review Board (ITRB) per GISTM 2020 Requirement 14 before resuming AI-driven phreatic surface management. See also: geotechnical slope monitoring AI prompt injection (related embankment deformation AI adversarial surface) and dam safety monitoring AI prompt injection (related dam monitoring AI regulatory gap context). Get early access

Related questions