Prompt injection in mining and mineral processing AI

Four adversarial injection surfaces in mining and mineral processing AI

1. Drill core sample image AI injection (Hexagon Mining AI, ABB Ability Genix AI, CoreScan AI)

Drill core sample AI processes photographs of drill core samples, core tray images, and lithogeochemical logging photographs submitted through AI-assisted geological interpretation platforms that extract mineral grade estimates, lithological classifications, and ore body boundary determinations from these image inputs, generating ore reserve grade estimates that are incorporated into NI 43-101 (National Instrument 43-101, Canadian Securities Administrators) and JORC (Joint Ore Reserves Committee, Australasian) compliant mineral resource and reserve reports filed with securities regulators and used as the primary basis for mining project capital valuations on the TSX, ASX, and major mining equity markets. Hexagon Mining AI is deployed across major exploration and operating mine projects where AI-assisted core analysis tools process drill core sample photographs to generate grade estimates incorporated into resource model updates, mine plan revisions, and feasibility study documents that become the basis for public mineral resource disclosures. ABB Ability Genix AI for industrial analytics processes geological data display images and core sample photograph submissions from mine sites where ABB’s AI tools generate grade control and resource estimation outputs for mine planning decisions. CoreScan AI, deployed by major mining companies and independent geological consulting firms, processes drill core sample tray photographs through AI-assisted automated core scanning and mineralogy classification tools that generate mineral abundance data incorporated into resource estimation workflows for NI 43-101 and JORC resource reports.

The adversarial injection surface is the drill core sample photograph and core tray image submission pathway: photographs of drill core samples laid in core trays, close-up images of core sample mineral zones, and lithogeochemical logging photographs submitted by geologists or automated core scanning systems to Hexagon Mining AI, ABB Ability Genix AI, or CoreScan AI for AI grade estimation and mineralogical classification. An adversarially crafted drill core sample photograph — in which pixel perturbations applied to the mineral grain visible region, colour intensity values corresponding to high-grade mineralisation, or textural features associated with ore-bearing zones cause the AI to extract an inflated mineral grade estimate — can cause the AI to over-report the ore grade for the affected core interval, inflating the mineral resource estimate incorporated into the NI 43-101 or JORC resource report filed with the TSX or ASX. The scale of financial impact from adversarially inflated drill core grade estimates in mining resource reports is exceptional: a 10% inflation in ore grade estimate for a 100 million tonne mineral resource can represent hundreds of millions of dollars in inflated project value on a mining equity, and NI 43-101-compliant resource estimates are the primary basis on which mining equity capital raises and M&A transactions are priced.

The regulatory consequences of adversarially falsified mineral grade estimates through drill core sample AI injection are severe under both mining securities and criminal law. NI 43-101 (TSX/CSE/VSX) and JORC (ASX) impose strict disclosure obligations on mining companies filing mineral resource and reserve estimates with securities regulators, requiring that disclosed resource estimates are supported by competent person reports that accurately reflect the underlying geological data — resource estimates that incorporate adversarially inflated AI-generated grade data from core sample photographs are false securities disclosures subject to CSA (Canadian Securities Administrators) and ASIC (Australian Securities and Investments Commission) enforcement, with potential criminal securities fraud liability and disgorgement of inflated capital raise proceeds. The Québec Securities Act and Ontario Securities Act impose civil liability for false and misleading disclosure in mineral resource reports filed with Canadian regulators; US secondary market fraud liability under SEC Rule 10b-5 (17 CFR 240.10b-5) attaches where adversarially inflated resource reports are relied upon by US investors in cross-listed mining securities. ASX/TSX mining securities liability exposure for adversarially inflated resource reports includes stock exchange delisting, regulatory enforcement actions, and class action exposure from investors who acquired shares on the basis of the falsified resource estimate. Threshold: 55 for drill core sample image AI.

2. Ore sorting conveyor belt camera AI injection (Tomra AI, Steinert AI, Sor-Sense AI)

Ore sorting AI processes real-time conveyor belt camera images, ore particle photograph streams, and mineral abundance display images submitted through AI-assisted ore sorting platforms that classify ore particles by mineral grade, lithological type, and grade cut-off threshold as the ore passes through the sorter machine on a conveyor belt, generating sorting decisions — accept or reject — for each particle or batch that determine which material is directed to the processing plant and which is directed to the waste stream. Tomra Mining AI is deployed at ore sorting installations worldwide, including copper, gold, diamond, and battery metal mining operations, processing real-time conveyor camera images of ore particles through AI-assisted sorting algorithms that classify each particle against a grade threshold and direct a compressed air ejector to remove below-threshold particles from the ore stream before it enters the processing plant. Steinert AI ore sorting systems, deployed at zinc, tin, copper, and fluorite mining operations across Europe, Australia, and the Americas, process real-time XRT (X-Ray Transmission) and NIR (Near-Infrared) sensor image streams of ore particles through AI grade classification tools that generate per-particle sorting decisions at throughput rates of hundreds of tonnes per hour. Sor-Sense AI ore sorting technology, deployed at precious metals and base metals mining operations, processes camera image feeds of ore particle streams through AI mineral abundance classification tools that generate sorting decisions for open-pit and underground mine run-of-mine ore.

The adversarial injection surface is the conveyor belt camera image feed, ore particle photograph stream, and mineral abundance display image submission pathway: real-time camera images of ore particles on the sorter conveyor belt, XRT sensor image streams of ore particle cross-sections, and NIR image feeds of ore particle mineral abundance submitted to Tomra AI, Steinert AI, or Sor-Sense AI for AI grade classification and sort decision generation. An adversarially crafted conveyor belt ore particle camera image — in which pixel perturbations applied to the mineral grain pattern, colour saturation values corresponding to high-grade mineralisation, or XRT density pattern of the ore particle cause the Tomra AI or Steinert AI to classify a high-grade ore particle as below the cut-off threshold and direct it to the waste stream — causes the AI sorter to reject valuable ore particles, reducing mineral recovery in the processing plant, inflating waste tonnage, and reducing the mine’s overall mineral recovery rate. The inverse attack — causing low-grade waste particles to be classified as above-threshold ore and directed to the processing plant — dilutes the ore feed grade to the processing plant, reducing processing efficiency and increasing energy consumption per tonne of recovered metal.

The regulatory and operational consequences of adversarially manipulated ore sorting AI classifications span mineral resource recovery obligations, securities disclosure accuracy, and MSHA worker safety dimensions. Mine operator agreements with project financiers and offtake purchasers typically include minimum ore recovery warranties and concentrate quality specifications; adversarial manipulation of ore sorting AI that systematically misclassifies ore grade causes contract performance failures under offtake agreements that specify minimum payable metal content in concentrate shipments. For mines operating under royalty agreements or streaming arrangements, adversarially reduced ore recovery that inflates waste tonnage and reduces reported mill feed grade affects royalty and stream payment calculations — creating royalty fraud exposure where the manipulation is attributable to a party with a royalty or stream payment obligation. MSHA 30 CFR Part 56 (Safety and Health Standards — Surface Metal and Nonmetal Mines) imposes maintenance and equipment safety standards for ore sorting conveyor systems; adversarial manipulation of ore sorting AI that affects equipment condition monitoring data display images submitted to maintenance AI tools creates MSHA 30 CFR Part 50 injury reporting compliance exposure. Threshold: 55 for ore sorting conveyor belt camera AI.

3. Equipment condition photograph AI injection (Komatsu FrontRunner AI, Caterpillar MineStar AI, Hitachi AI)

Mining equipment condition AI processes haul truck inspection photographs, dragline and shovel structural condition images, conveyor belt condition survey photographs, and mining equipment onboard sensor display images submitted through AI-assisted fleet management and predictive maintenance platforms that extract equipment condition classifications, component wear assessments, and maintenance action recommendations from these image inputs, generating maintenance schedules and equipment health reports that determine whether mining equipment is approved for continued operations, mandated for scheduled maintenance, or classified as requiring immediate removal from service under MSHA safety compliance requirements. Komatsu FrontRunner (Modular Mining) AI is deployed across autonomous haulage fleets at Rio Tinto’s Pilbara operations, BHP’s Pilbara and South Flank operations, and major copper mines in Chile and Peru, processing haul truck onboard condition monitoring display photographs and equipment health sensor image submissions through AI-assisted fleet management tools that generate truck maintenance schedules and safety compliance reports for MSHA and national mining regulator review. Caterpillar MineStar AI is deployed across Cat autonomous and semi-autonomous haul truck fleets at Rio Tinto, BHP, Codelco, and other Tier 1 mining operators, processing truck frame condition photographs, tyre wear survey images, and haul truck structural inspection photographs through AI equipment health management tools that generate maintenance recommendations and MSHA-required safety equipment inspection records. Hitachi EX Mining AI processes shovel and excavator condition inspection photographs through AI-assisted predictive maintenance tools deployed at major iron ore, coal, and copper mining operations.

The adversarial injection surface is the haul truck inspection photograph, mining equipment structural condition image, and onboard sensor display screenshot submission pathway: photographs of haul truck frame welds, tyre condition surveys, dragline structural inspection images, conveyor belt splice and idler condition photographs, and equipment health monitoring display screenshots submitted by mine site maintenance technicians or automated inspection systems to Komatsu FrontRunner AI, Caterpillar MineStar AI, or Hitachi AI tools for condition assessment and maintenance recommendation generation. An adversarially crafted haul truck inspection photograph — in which pixel perturbations applied to the frame crack indication, tyre sidewall condition region, or structural weld image area cause the Komatsu AI or MineStar AI to classify the equipment condition as within acceptable service limits when the unperturbed photograph would trigger an immediate remove-from-service recommendation under MSHA-required equipment safety standards — can suppress a mandatory safety maintenance action that would otherwise remove the defective haul truck from operations. The safety consequence is that a haul truck with an undetected structural defect continues in autonomous haulage operations, creating MSHA 30 USC § 820 civil penalty exposure of $72,530 per knowing violation for the mine operator and potential criminal penalty exposure for supervisory personnel who authorised continued operations.

The regulatory consequences of adversarially suppressed equipment safety defect detection in mining AI are governed by MSHA’s civil and criminal enforcement authority under the Federal Mine Safety and Health Act 1977 (30 USC § 801 et seq.). MSHA 30 CFR Part 56 (Surface Metal and Nonmetal Mines) and Part 57 (Underground Metal and Nonmetal Mines) impose equipment safety and maintenance inspection requirements for mining equipment, with mandatory inspection intervals, defect reporting obligations, and removal-from-service standards for defective equipment — failure to comply with equipment inspection requirements is a civil violation under 30 USC § 820, with maximum civil penalties of $72,530 per knowing violation and mandatory minimum penalties for significant and substantial (S&S) violations. Knowing or wilful violations of MSHA equipment safety requirements that contribute to a fatal or serious mine accident create criminal liability under 30 USC § 820(d), with imprisonment sentences of up to five years for wilful MSHA violations resulting in death. MSHA’s Pattern of Violations (POV) programme can result in mine closure orders where a mine operator accumulates a pattern of equipment safety violations — adversarial manipulation of equipment inspection AI that systematically suppresses defect flags for safety-significant equipment conditions creates POV programme exposure for the mine operator. Threshold: 55 for equipment condition photograph AI.

4. Environmental monitoring photograph AI injection (Hexagon Mining AI, Intelex AI, Cority AI)

Mining environmental monitoring AI processes tailings storage facility (TSF) condition photographs, acid mine drainage (AMD) monitoring site images, surface water quality monitoring station photographs, and mine reclamation progress survey images submitted through AI-assisted environmental compliance management platforms that extract environmental compliance status classifications, exceedance indicators, and reclamation completion assessments from these image inputs, generating environmental compliance records that are submitted to state and federal regulators under NPDES permit reporting requirements, CWA Section 319 nonpoint source management program reporting obligations, and mine reclamation bond release applications. Hexagon Mining AI processes TSF condition survey photographs and AMD monitoring site images through AI-assisted environmental monitoring and compliance tools deployed at major open-pit mining operations, generating TSF condition assessments and AMD exceedance classifications that form part of the mine operator’s NPDES permit compliance reporting. Intelex AI, deployed at Fortune 500 mining and industrial operations, processes environmental monitoring station photographs and compliance inspection images through AI-assisted EHS compliance management tools that generate NPDES permit reporting data and environmental compliance records for regulatory submission. Cority AI processes environmental monitoring data display photographs and compliance survey images through AI-assisted EHS management tools for multinational mining operators, generating environmental compliance reports submitted under CWA, CERCLA, and state mining reclamation regulatory frameworks.

The adversarial injection surface is the tailings facility condition photograph, acid mine drainage monitoring site image, and surface water quality monitoring station photograph submission pathway: photographs of tailings storage facility embankments and spillways, AMD seep monitoring site images, water quality sampling station photographs, and mine reclamation progress survey images submitted to Hexagon Mining AI, Intelex AI, or Cority AI for AI environmental compliance classification and regulatory reporting data generation. An adversarially crafted tailings facility condition photograph — in which pixel perturbations applied to the embankment seepage indicator region, spillway erosion area, or phreatic surface indicator zone cause the AI to classify the TSF condition as compliant with regulatory minimum stability and seepage standards when the unperturbed photograph would indicate a stability concern requiring regulatory notification — can suppress an environmental compliance flag that would otherwise trigger mandatory regulatory reporting under state mining permit conditions and MSHA TSF safety regulations, allowing a potentially unstable tailings facility to continue operating without required regulatory intervention.

The regulatory consequences of adversarially suppressed environmental compliance defect detection in mining AI span CWA, CERCLA, and state mining reclamation law dimensions. CWA Section 309 (33 USC § 1319) imposes civil penalties of up to $25,000 per day per violation for NPDES permit exceedances that are not reported as required — adversarial suppression of an AI environmental monitoring exceedance flag that prevents required NPDES reporting creates CWA § 309 civil penalty exposure from the date the unreported exceedance occurred. CWA Section 319 nonpoint source management program requirements impose reporting obligations for nonpoint source pollution incidents at mining operations; AMD exceedances at mines that discharge to designated CWA § 319 watershed management areas carry additional state regulatory consequences. CERCLA liability attaches to operators of facilities from which hazardous substances are released — tailings storage facility failures that result in hazardous substance releases carry CERCLA Superfund cleanup cost recovery liability against the mine operator and its parent company, with cleanup costs at major TSF failure incidents exceeding $100 million. State mining reclamation bond regulations in major mining states — Nevada NRS Chapter 519A, Arizona ARS Title 27, Colorado CRS Title 34 — require mine operators to post reclamation bonds sized to cover full mine reclamation costs; adversarial environmental monitoring AI suppression that conceals reclamation progress shortfalls from state regulators creates bond forfeiture risk at reclamation completion review. Threshold: 50 for environmental monitoring photograph AI, reflecting Superfund liability severity.

Integration: mining and mineral processing AI image ingestion with Glyphward pre-scan

Mining and mineral processing AI image ingestion flows from core sample photograph portals, ore sorting camera feeds, equipment inspection image channels, and environmental monitoring photograph submissions into geological interpretation AI, ore sorting AI, equipment health AI, and environmental compliance AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary — particularly for externally sourced core sample photographs, third-party sorter camera feeds, maintenance technician equipment inspection submissions, and contractor-submitted environmental monitoring images — before AI-generated output is committed to resource reports, sorting records, maintenance schedules, or environmental compliance filings:

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"

# Mining and mineral processing AI — ore grade inflation in NI 43-101/JORC
# securities disclosures, ore sorter misclassification, MSHA equipment safety
# flag suppression, CWA/CERCLA environmental exceedance concealment.
# MSHA 30 CFR Parts 50-57, NI 43-101/JORC, CWA §319, CERCLA,
# state mining reclamation bond regulations.
THRESHOLD_MINING_ENV     = 50   # environmental monitoring (CERCLA severity)
THRESHOLD_MINING_DEFAULT = 55   # core sample, ore sorting, equipment condition


class MiningAIContext(str, Enum):
    DRILL_CORE_SAMPLE   = "drill_core_sample"   # Hexagon Mining, ABB Genix, CoreScan
    ORE_SORTING         = "ore_sorting"         # Tomra, Steinert, Sor-Sense
    EQUIPMENT_CONDITION = "equipment_condition" # Komatsu FrontRunner, Cat MineStar, Hitachi
    ENVIRONMENTAL       = "environmental"       # Hexagon Mining, Intelex, Cority


def threshold_for(context: MiningAIContext) -> int:
    if context == MiningAIContext.ENVIRONMENTAL:
        return THRESHOLD_MINING_ENV
    return THRESHOLD_MINING_DEFAULT


async def scan_mining_image(
    image_path: str | Path,
    context: MiningAIContext,
    facility_id_hash: str,  # SHA-256 of mine site / facility identifier
    batch_id: str,          # e.g. drill hole ID, truck ID, TSF zone, sorter line
    work_ref: str,          # e.g. "WO-2026-44721", "NPDES-UT0026006-Q2-2026"
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan a mining or mineral processing AI image for adversarial injection
    payloads before forwarding to a drill core sample AI, ore sorting AI,
    equipment condition AI, or environmental monitoring AI.

    Raises AdversarialMiningImageError if the Glyphward score meets or
    exceeds the context-specific threshold (50 for environmental, 55 for others).
    """
    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": {
                "mining_context":    context.value,
                "facility_id_hash":  facility_id_hash,
                "batch_id":          batch_id,
                "work_ref":          work_ref,
                "client_scan_id":    client_scan_id,
                "image_sha256":      image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "facility_id_hash": facility_id_hash,
        "batch_id":         batch_id,
        "work_ref":         work_ref,
        "mining_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_mining_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialMiningImageError(
            f"Mining AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"facility={facility_id_hash} ref={work_ref}"
        )
    return result


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


class AdversarialMiningImageError(Exception):
    """Raised when a mining AI image exceeds the adversarial injection threshold."""
    pass

Call scan_mining_image() with MiningAIContext.DRILL_CORE_SAMPLE before forwarding core sample photographs to Hexagon Mining AI, ABB Ability Genix AI, or CoreScan AI — this is the highest-consequence integration for securities fraud exposure, where adversarially inflated grade estimates propagate into NI 43-101/JORC mineral resource disclosures on the TSX and ASX. Call with MiningAIContext.ORE_SORTING for conveyor belt camera image feeds before Tomra AI, Steinert AI, or Sor-Sense AI grade classification, using batch_id to track the specific sorter line and ore lot for operational audit and offtake contract compliance purposes. Call with MiningAIContext.EQUIPMENT_CONDITION for haul truck inspection and dragline condition photographs before Komatsu FrontRunner AI, Caterpillar MineStar AI, or Hitachi AI health management, with work_ref linking scan records to MSHA-required equipment safety inspection records. Call with MiningAIContext.ENVIRONMENTAL — threshold 50, strictest for CERCLA Superfund severity — for tailings facility photographs and AMD monitoring site images before Hexagon Mining AI, Intelex AI, or Cority AI environmental compliance classification, preserving image_sha256 for NPDES permit reporting audit trail reconstruction. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four mining AI injection surfaces; covers how adversarial pixel-level perturbations cause AI misclassification without detectable visual artifacts at standard review resolution.
• Vision-language model security — technical architecture of adversarial image attacks against vision-language models including the pixel perturbation classes applicable to core sample photograph grade inflation and ore sorting camera manipulation.
• Prompt injection scanner for document AI — document AI scanning covering the scanned document injection vectors applicable to NI 43-101 competent person report and MSHA inspection record manipulation.
• Water treatment and environmental AI prompt injection — environmental monitoring AI injection attacks with complementary CWA and CERCLA regulatory context applicable to tailings facility and AMD monitoring AI systems.
• Free tier — 10 scans/day, no card required — start scanning mining AI images at development volumes before committing to a production plan.