Prompt injection in mining and oil & gas AI

Four adversarial injection surfaces in mining and oil & gas AI

1. Drillcore and borehole image injection in geological AI (MineSense, Maptek Eureka AI, Trimble MineStar)

Geological AI platforms process photographic images of drillcore (cylindrical rock samples extracted from drill holes during exploration and production drilling) and drill chip samples (rock fragments produced by rotary or percussion drilling) to classify rock types, estimate mineral grades, identify alteration zones, and characterise ore-waste contacts — determinations that feed block models used for mineral resource and reserve estimates under JORC Code (Australia), NI 43-101 (Canada), and SEC Regulation S-K 1300 (United States) reporting standards. MineSense Technologies’ ShovelSense and BeltSense AI platforms use X-ray fluorescence (XRF) sensors combined with camera imaging at loading shovels and conveyor belts to grade ore-waste material in real time at copper, gold, and nickel mines operated by Teck Resources, BHP, and Glencore. Maptek’s Eureka geological modelling platform and PointStudio AI incorporate automated core photograph classification to assign lithology codes and grade estimates to core tray images submitted by geologists through Maptek’s data portal. The adversarial injection surface involves drillcore photograph submissions: core photographs are standard geological documentation submitted by third-party core logging contractors, exploration geologists, and laboratory analysts through geological data management systems (GEOVIA, Datamine, acQuire) as part of resource estimation campaigns. An adversarially crafted core tray photograph — in which pixel-level perturbations applied to the visible mineralisation patterns (pyrite dissemination, visible copper staining, quartz vein density indicators) cause the AI geological classifier to assign a higher-grade lithology code than the physical mineralogy supports — submitted through a geological data portal by a contracted core logging service can corrupt the block model feeding a mineral resource estimate. For publicly listed mining companies reporting JORC or NI 43-101 resource estimates, a systematically inflated AI-derived grade estimate creates a material misrepresentation risk under securities law — with consequences analogous to the Bre-X Minerals scandal (fraudulent core assays inflating reported gold grades) but executed through adversarial AI manipulation rather than physical salting. The converse attack — adversarially crafted images that cause the AI to under-report grade at a rival’s mineral deposit or in a valuation due diligence exercise — has corporate intelligence and M&A manipulation applications. Trimble MineStar’s fleet management and mine control AI also incorporates face profiling and blast design image classification from photographs of blast holes and mine faces; adversarial injection in this context can corrupt blast design AI that optimises drill pattern and explosive loading parameters — with process safety consequences if incorrect blast parameters are acted on.

2. Pipeline corrosion inspection AI injection (Baker Hughes AIQ, Schlumberger Delfi AI, ABB Asset Vision)

Pipeline integrity AI processes inspection images from inline inspection (ILI) tools (magnetic flux leakage, ultrasonic wall measurement), external inspection drones (Flyability Elios 3 for internal tank and vessel inspection, Percepto Sparrow for external above-ground pipeline corridors), and robotic crawlers (ROSEN Group ROSENpig, PII Pipeline Solutions) to classify corrosion severity, crack depth, and wall thickness loss against the engineering assessment thresholds in the pipeline’s integrity management plan. Baker Hughes AIQ’s pipeline inspection AI classifies corrosion indications in drone inspection photographs and ILI data visualisations submitted by pipeline operators through the Baker Hughes Leucipa AI platform and Cordant digital operations platform. Schlumberger’s Delfi Digital Subsurface Environment processes pipeline inspection images submitted by offshore operators through operator data rooms and cloud-based integrity management portals. The adversarial injection surface involves inspection drone image submissions: pipeline inspection drone operators at midstream pipeline operators (Williams Companies, TC Energy, Enterprise Products Partners, ADNOC Pipelines) transfer inspection image sets from drone missions through secure upload portals to the pipeline integrity AI platform for automated defect classification. An adversarially crafted external corrosion inspection image — in which pixel-level perturbations applied to a corroded pipe surface area reduce the apparent pitting depth or surface area of an active corrosion feature — submitted through the inspection drone upload portal can cause the pipeline integrity AI to classify an active corrosion indication below the operator’s dig-and-repair trigger threshold, deferring a maintenance excavation action. For high-pressure gas transmission pipelines (class 3 and 4 locations, HCAs — High Consequence Areas) where a corrosion-driven pipeline failure results in a high-pressure gas rupture and ignition, the consequence of a deferred corrosion repair triggered by a corrupted AI inspection classification can be a fatal pipeline explosion — the category of event associated with incidents including the San Bruno PG&E pipeline explosion (2010, 8 deaths) and the Bellingham, Washington Olympic Pipe Line rupture (1999, 3 deaths). PHMSA’s pipeline integrity rule (49 CFR Part 192 and Part 195) requires operators to use engineering-critical assessment (ECA) methods for integrity decisions; adversarial manipulation of the inspection AI input data that feeds ECA calculations constitutes corruption of the integrity management process required by federal pipeline safety regulation.

3. Wellsite operations camera AI injection (TotalEnergies AI, Weatherford AI, Keyera)

Wellsite operations AI processes camera images from production well pads, compression stations, and processing facilities to monitor equipment health, detect safety hazards (hydrocarbon leaks, flare status, personnel proximity to hazardous equipment), and assess production equipment condition. TotalEnergies’ AI-powered wellsite monitoring platform, deployed at upstream production assets in the North Sea, Middle East, and North Africa, uses camera feeds and periodic site photograph submissions to classify equipment condition — surface wellhead equipment, Christmas tree valve status, separator condition, and flare system status — through AI models integrated with the TotalEnergies remote operations centre. Weatherford AI and Keyera’s gas processing AI platforms incorporate wellsite equipment photograph submissions from field operators and third-party field inspection contractors through web portal and mobile application upload workflows to support AI-assisted equipment health assessment and work order generation in CMMS (Computerised Maintenance Management Systems) platforms. The adversarial injection surface involves wellsite inspection photograph submissions: field operators at remote well pad locations and third-party inspection contractors routinely submit wellsite equipment condition photographs through mobile applications (Inpixon, Field360, iSite) and operator portal upload forms as part of well pad inspection routines and scheduled maintenance programmes. An adversarially crafted wellsite equipment photograph — in which pixel-level perturbations applied to a surface safety valve (SSV), emergency shutdown valve (ESDV), or wellhead pressure gauge image cause the AI equipment condition classifier to assign a “normal operating condition” status when the physical equipment shows a visible leak path, valve stem corrosion, or pressure gauge damage — submitted through a mobile field inspection application can suppress a work order generation that the AI would otherwise have triggered. For upstream oil and gas production facilities where a failed surface safety valve or blowout preventer component reaches a critical failure state without triggering a maintenance action, the consequence is an uncontrolled well kick or blowout — the category of incident associated with the Macondo Deepwater Horizon disaster (2010) and the Piper Alpha platform explosion (1988, 167 deaths). The adversarial injection surface is accessible to a field contractor who has legitimate access to the mobile inspection application but wants to suppress a maintenance action for financial reasons (deferred work order = deferred shutdown cost).

4. Offshore platform inspection drone image injection (Cognite AI, ABB Asset Vision, Aker Solutions IRM)

Offshore structural integrity AI processes inspection images from underwater ROVs, above-water aerial drones, and close-visual inspection (CVI) teams submitted through offshore integrity management platforms to classify structural defects, corrosion severity, coating degradation, and fatigue crack indicators on production platforms, FPSOs, and subsea structures. Cognite Data Fusion’s integrity management module, deployed by Equinor, Aker BP, and Vår Energi on Norwegian Continental Shelf (NCS) platforms, processes ROV inspection video frames and drone inspection images submitted by inspection service providers (Stinger Technology, Oceaneering International, Subsea 7 IMR) through Cognite’s API and data connector integrations. Aker Solutions’ IRM AI platform and ABB Ability Asset Vision process offshore structure inspection images through integrity management portals used by offshore operators across the UK Continental Shelf (UKCS), Gulf of Mexico, and deepwater West Africa. The adversarial injection surface involves inspection service provider image submissions: third-party inspection service providers — Oceaneering International, TechnipFMC, Saipem — transfer ROV inspection video frames, multi-beam sonar images, and close-visual inspection photographs to offshore operators through data transfer portals, cloud storage integrations (Microsoft Azure, AWS S3), and direct API uploads. An adversarially crafted ROV inspection image — in which pixel-level perturbations applied to a jacket leg node, a conductor guide frame, or a mooring chain inspection area cause the AI structural classification model to assign a “within threshold — monitor next inspection cycle” classification when the physical feature exceeds the crack acceptance criteria in the structural integrity management plan (SIMP) — submitted through a cloud portal data transfer can defer a repair action that the AI would otherwise have flagged for immediate intervention. For North Sea and Gulf of Mexico fixed platforms in the late-life phase — where fatigue crack growth rates in jacket legs and conductor guides are accelerating — a deferred structural repair triggered by a corrupted AI classification contributes to the structural failure risk profile. HSE (UK) Offshore Installations (Prevention of Fire and Explosion, and Emergency Response) Regulations 1995 (PFEER) and BSEE (Bureau of Safety and Environmental Enforcement) structural integrity requirements establish the legal framework for offshore structural integrity management; adversarial corruption of the AI inspection data layer that feeds structural integrity assessments creates regulatory liability analogous to falsified inspection records under these frameworks.

Integration: mining and oil & gas AI image ingestion with Glyphward pre-scan

Mining and oil & gas AI image ingestion flows from geological data portal uploads, pipeline inspection drone transfer portals, wellsite mobile inspection applications, and offshore integrity management cloud integrations into AI processing queues. Insert Glyphward’s pre-scan at the ingestion boundary before images reach the geological, pipeline integrity, or structural AI:

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"

# Process safety threshold: pipeline integrity failure or structural collapse
# has spill/explosion/fatality consequence. 55 for standard; 50 for safety-critical.
THRESHOLD_MINING_ENERGY_AI = 55
THRESHOLD_PROCESS_SAFETY_CRITICAL = 50  # surface safety valves, SSVs, BOP systems


class MiningEnergyAIContext(str, Enum):
    GEOLOGICAL_CORE = "geological_core"             # drillcore / drill chip images
    PIPELINE_INSPECTION = "pipeline_inspection"     # corrosion / ILI / drone images
    WELLSITE_EQUIPMENT = "wellsite_equipment"       # surface equipment condition photos
    OFFSHORE_STRUCTURAL = "offshore_structural"     # platform / FPSO structural inspection


async def scan_energy_image(
    image_source: str | Path | bytes,
    context: MiningEnergyAIContext,
    site_id_hash: str,         # SHA-256 of site/asset identifier (not raw)
    asset_id_hash: str,        # SHA-256 of component/asset identifier (not raw)
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan a mining / oil & gas AI image for adversarial injection payloads before
    forwarding to geological, pipeline integrity, or structural AI.
    """
    if isinstance(image_source, (str, Path)):
        image_bytes = Path(image_source).read_bytes()
    else:
        image_bytes = image_source

    image_b64 = base64.b64encode(image_bytes).decode()
    image_sha256 = hashlib.sha256(image_bytes).hexdigest()
    scan_id = str(uuid.uuid4())

    threshold = (
        THRESHOLD_PROCESS_SAFETY_CRITICAL
        if context == MiningEnergyAIContext.WELLSITE_EQUIPMENT
        else THRESHOLD_MINING_ENERGY_AI
    )

    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json={
            "image": image_b64,
            "source": context.value,
            "metadata": {
                "energy_context": context.value,
                "site_id_hash": site_id_hash,
                "asset_id_hash": asset_id_hash,
                "client_scan_id": scan_id,
                "image_sha256": image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "site_id_hash": site_id_hash,
        "asset_id_hash": asset_id_hash,
        "energy_context": context.value,
        "scan_id": result["scan_id"],
        "client_scan_id": 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_energy_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialEnergyImageError(
            f"Mining/energy AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"site={site_id_hash[:8]}... asset={asset_id_hash[:8]}..."
        )
    return result


async def scan_inspection_mission(
    image_paths: list[Path],
    context: MiningEnergyAIContext,
    site_id_hash: str,
    asset_id_hash: str,
) -> dict:
    """Scan a drone inspection mission batch concurrently before AI processing."""
    allowed, blocked, errors = [], [], []

    async with httpx.AsyncClient() as client:
        tasks = [
            scan_energy_image(p, context, site_id_hash, asset_id_hash, client)
            for p in image_paths
        ]
        results = await asyncio.gather(*tasks, return_exceptions=True)

    for path, result in zip(image_paths, results):
        if isinstance(result, AdversarialEnergyImageError):
            blocked.append({"path": str(path), "error": str(result)})
        elif isinstance(result, Exception):
            errors.append({"path": str(path), "error": str(result)})
        else:
            allowed.append({"path": str(path), "scan_id": result["scan_id"]})

    return {
        "site_id_hash": site_id_hash,
        "context": context.value,
        "total": len(image_paths),
        "allowed": len(allowed),
        "blocked": len(blocked),
        "errors": len(errors),
        "blocked_items": blocked,
    }


async def write_energy_audit_record(record: dict) -> None:
    """Persist audit record to your integrity management audit log (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


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

The site_id_hash and asset_id_hash fields provide the integrity management evidence chain: a blocked inspection image record links scan_id + site_id_hash + asset_id_hash + image_sha256 for post-incident investigation without storing raw facility or asset identifiers in the audit log. For pipeline and offshore structural inspection contexts, route blocked images to an immediate integrity alert — a deferred inspection action triggered by a blocked adversarial image should default to the conservative physical inspection option rather than relying on the AI classification. Get early access

Coverage matrix

Related questions

Further reading

• Energy and utilities AI adversarial images — adjacent infrastructure inspection AI surface covering power grid transmission inspection, solar panel defect AI, and wind turbine blade inspection with overlapping drone inspection attack vectors.
• Manufacturing quality inspection AI prompt injection — related quality inspection AI surface covering industrial inspection AI including processing equipment and asset integrity adjacent to oil and gas operations.
• Indirect prompt injection via image — foundational attack pattern covering adversarial pixel perturbation delivery through legitimate submission pathways applicable to all mining and energy AI surfaces.
• Supply chain and logistics AI prompt injection — related supply chain AI surface covering commodity transport and logistics AI adjacent to mining and oil & gas distribution.
• Free tier — 10 scans/day, no card required — start scanning mining and energy AI images at development volumes before committing to a production plan.