Prompt injection in pipeline integrity inspection AI

Four adversarial injection surfaces in pipeline integrity inspection AI

1. MFL anomaly map image AI (Baker Hughes FlexPIG, TDW SmartScan, NDT Global)

Magnetic flux leakage (MFL) inspection pigs magnetise the pipe wall to near-saturation using permanent magnets and measure flux leakage at each discrete sample position using Hall-effect sensor arrays. The raw sensor data is rendered into an MFL anomaly map — a 2D colour-coded image where the horizontal axis represents the axial pipe position (in metres from a reference launcher), the vertical axis represents the circumferential sensor array position (clock position from 12 o’clock), and the pixel colour encodes flux leakage amplitude (red = high leakage = metal loss; blue = background). AI classification of MFL anomaly map images determines: whether a candidate flux leakage signal is a real metal-loss defect (corrosion pit, surface gouge) or a benign feature (weld seam, mill scale variation, circumferential weld cap); the defect depth (percentage of nominal wall thickness affected — the primary input to remaining strength calculations under ASME B31G or RSTRENG); the defect length and width in the axial and circumferential planes; and whether the signal matches patterns characteristic of stress corrosion cracking (SCC) — which produces a distinctive MFL signal profile of closely spaced low-amplitude anomalies that an MFL AI trained on SCC colonies identifies as distinct from corrosion pitting.

Baker Hughes’ FlexPIG AI classifier processes rendered MFL anomaly map TIFF images from the pig’s 1,000+ sensor channel output at a resolution of 1mm axial × 5mm circumferential per pixel. TDW’s SmartScan AI and NDT Global’s AI classification system process architecturally equivalent renders — the MFL anomaly map image format is standardized across ILI vendors under the Pipeline Operators Forum (POF) ILI data specification, meaning the adversarial injection surface is structurally identical across all major ILI providers. An adversarial perturbation on a rendered MFL anomaly map image that reduces the peak red-channel intensity of a corrosion pit signature — shifting the anomaly from 35% wall loss (above the ASME B31G 80% SMYS burst pressure calculation threshold for repair) to 25% wall loss (below threshold, “monitor only”) — causes the ILI AI to classify the defect as “no immediate action required” in the fitness-for-service (FFS) assessment, bypassing excavation and repair. In a high-consequence area (HCA) crossing within 200 metres of a population centre, the operator’s PHMSA-required 6-month remediation deadline for critical anomalies (49 CFR 195.452(h)(4)) is never triggered because the AI-generated defect call report does not flag the anomaly as exceeding the repair threshold.

2. UT B-scan cross-sectional image AI (ROSEN RoCorr UT-C, Eddyfi Technologies, Olympus NDT)

Ultrasonic testing (UT) ILI pigs transmit pulsed ultrasonic beams perpendicular to the pipe wall and measure the time-of-flight of the return pulse from the inner and outer pipe surface to compute wall thickness at each measurement point. The UT data is rendered into B-scan images — cross-sectional views of the pipe wall where the horizontal axis represents the axial pipe position and the vertical axis represents the wall thickness from ID to OD, with pixel intensity encoding the UT amplitude response at each depth. AI classification of UT B-scan images determines whether anomalies in the B-scan represent: general corrosion (broad wall-thinning zones, ID or OD surface origin); pitting corrosion (localised sharp-walled depressions, higher depth-to-width ratio than general corrosion); axial cracks (linear anomalies oriented parallel to the pipe axis — EMAT UT is the preferred tool but conventional UT detects large cracks); circumferential cracks (linear anomalies perpendicular to flow, detected by shear-wave UT); laminar flaws (hydrogen-induced cracking, anomalies at mid-wall depth rather than surface-origin). ROSEN’s RoCorr UT-C AI and Eddyfi Technologies’ UT AI classifiers process rendered UT B-scan JPEG or TIFF images from the pig’s circumferential transducer array, generating defect call reports in PHMSA-required ILI data formats (POF PODS 4.0 or operator-specific defect CSV).

An adversarial perturbation on a rendered UT B-scan image that introduces a smoothing artefact across the depth profile of a hydrogen-induced cracking (HIC) laminar flaw — blurring the mid-wall amplitude signature into the surrounding material noise floor — causes the UT AI to classify the B-scan segment as “sound metal” rather than “HIC colony.” HIC — which forms in pipe manufactured from high-sulphur steel in wet hydrogen sulphide (H₂S) service environments (common in sour crude oil and produced water pipelines) — can propagate slowly for years before causing a step-change brittle fracture with no macroscopically visible deformation. A missed HIC colony in a sour gas transmission pipeline — operating at 8 MPa and transporting H₂S concentrations above the OSHA IDLH threshold of 50 ppm — that ruptures in an HCA produces a toxic vapour cloud with a potential lethal-concentration zone (IDLH within 1 km downwind under ALOHA dispersion modelling) and a secondary explosion risk from the natural gas component. The 2011 PG&E San Bruno rupture (operated at 69 bar; SCC + pre-existing manufacturing defect in seam weld not detected by ILI programme) established that ILI defect call report quality is a direct proximate cause of pipeline rupture consequences.

3. Aerial drone corrosion and coating inspection AI (AeroVironment HAPS, Percepto Arc, Flyability Elios)

Above-ground corrosion inspection of pipeline rights-of-way, valve stations, compressor stations, and above-ground facility piping uses autonomous drone patrol systems equipped with RGB and infrared cameras to detect external corrosion, coating disbondment, and mechanical damage. Percepto Arc autonomous drone stations (deployed at Chevron, ExxonMobil, and BP facility sites), Flyability Elios confined-space inspection drones, and AeroVironment HAPS (High Altitude Pseudo-Satellite) airborne inspection platforms process RGB camera frames and infrared thermal images through AI corrosion classifiers that identify: pipeline coating holidays (areas of disbonded or missing coating where the bare steel substrate is visible as a colour contrast against intact coating); rust staining patterns (surface corrosion indicating early-stage metal loss on exposed steel); external pitting corrosion (the three-dimensional pitting texture visible in close-range RGB images); and thermal anomalies in insulated piping (hot or cold spots indicating insulation failure and potential CUI — Corrosion Under Insulation — condensation zones that drive external corrosion beneath the insulation blanket). The AI classifier processes rendered drone camera frames at 4K resolution (3840×2160 pixels per frame) and generates inspection reports flagging GPS-tagged anomaly locations for follow-up excavation or coating repair scheduling.

An adversarial perturbation on a rendered drone RGB camera frame that shifts the colour signature of a coating holiday — filling in the exposed bare steel brownish tone with the olive-green texture of intact epoxy coating using a ±8 DN pixel shift within the JPEG compression noise floor — causes the corrosion AI to classify the frame as “coating intact, no anomaly” for a section of pipe that has a coating holiday and active external corrosion initiation site. External corrosion detected by drone patrol that exceeds the 20% wall loss threshold under NACE SP0169-2013 (“Control of External Corrosion on Underground or Submerged Metallic Piping Systems”) would trigger a cathodic protection survey and potential excavation; adversarial suppression of the coating holiday detection prevents this trigger. For onshore liquid petroleum pipelines operating in high-consequence areas, PHMSA’s atmospheric corrosion inspection requirements (49 CFR 195.583) require annual inspection of exposed piping and surface facilities — inspection intervals that depend entirely on the fidelity of the drone camera AI classification output for anomaly rate estimation.

4. ILI pig high-resolution camera image AI (CREAFORM PipeCheck, Eddyfi Pipe Scanner, GE Inspection Robotics)

High-resolution optical camera ILI systems — deployed in gas transmission pipelines where the medium is transparent (unlike liquid-filled crude oil lines) — transmit a camera pig through the pipe bore and record continuous high-definition video of the pipe interior surface. AI classification of ILI camera frame sequences determines: dent severity and shape (smooth dent vs. gouged dent vs. dent with crack — the crack-in-dent combination is a primary SCC initiation site); manufacturing defect identification (longitudinal seam weld quality, surface lamination, rolled-in scale from the original steel manufacturing process); mechanical damage classification (third-party interference — e.g., a backhoe impact creating a fresh gouge on the outer pipe surface visible as a circumferential bright scratch on the camera image — that triggers PHMSA’s 72-hour emergency excavation requirement under 49 CFR 192.613 / 195.402(c)(10)); and geometry anomalies (ovality, wrinkling, buckle formation — which affect pig travel and may indicate external soil loading or landslide movement on a geohazard corridor). Creaform PipeCheck AI, GE Inspection Robotics’ camera pig AI, and Eddyfi’s optical inspection AI process individual camera frames at 1920×1080 resolution using CNNs trained on labelled defect libraries from vendor inspection histories.

An adversarial perturbation on a rendered ILI camera frame that suppresses a fresh mechanical gouge signature — smoothing the bright linear scratch of a backhoe strike against the existing surface texture noise of the pipe bore — causes the camera AI to classify the frame segment as “clean bore, no mechanical damage” rather than triggering the PHMSA 72-hour emergency excavation requirement. Unreported mechanical damage (third-party interference) is a leading cause of pipeline failure in populated areas: the National Transportation Safety Board (NTSB) identified third-party damage as the leading cause of fatal natural gas pipeline accidents in the period 2010–2020 (NTSB SS-21-01, “Safety Study on Natural Gas Distribution Pipeline Integrity”). PHMSA’s damage prevention programme (49 CFR 192.614) requires operators to investigate third-party contact reports within 1 hour if the pipeline is in a class 3 or 4 location (10+ structures within 200m of centreline) — a response timeline that the ILI camera AI’s defect call report directly drives when a pig run is in progress.

Integration: pipeline integrity inspection AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for pipeline integrity AI belongs at the rendered image ingestion boundary before each AI classification step — before MFL anomaly map AI processes rendered heatmap images, before UT B-scan AI processes cross-sectional renders, before drone corrosion AI processes RGB camera frames, and before ILI camera AI processes pig video frames. Threshold 40 for pipeline integrity AI contexts reflects HCA consequence severity and the availability of complementary ILI technologies for cross-validation.

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"

# Pipeline integrity AI contexts: threshold 40
# PHMSA 49 CFR 195.452 / 192.917, API 1160, ASME B31G / RSTRENG.
PIPELINE_AI_THRESHOLD = 40


class PipelineAIContext(Enum):
    MFL_ANOMALY_MAP    = "mfl_anomaly_map"    # MFL pig rendered heatmap AI
    UT_BSCAN           = "ut_bscan"           # UT pig B-scan cross-section AI
    DRONE_CORROSION    = "drone_corrosion"    # Aerial RGB/IR coating inspection AI
    CAMERA_PIG         = "camera_pig"         # ILI optical camera pig frame AI


class AdversarialPipelineImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in a pipeline
    integrity AI rendered image above threshold 40.

    Consequence if not raised: corrosion pit / SCC / crack suppressed from
    defect call report → no excavation trigger → pipe segment remains in
    service past repair threshold → rupture in HCA → vapour cloud explosion
    or hazardous liquid spill with multi-kilometre consequence radius.
    Fail-safe: suppress AI classification, route to qualified ILI data analyst
    (per API 1160 Section 6.2.2 human review requirement) for manual review.
    """

    def __init__(self, scan_id: str, score: int,
                 context: PipelineAIContext,
                 pig_run_id: str, chainage_m: float | None,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.pig_run_id = pig_run_id
        self.chainage_m = chainage_m
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial pipeline image: "
            f"context={context.value} score={score} "
            f"pig_run={pig_run_id} chainage={chainage_m}m scan_id={scan_id}"
        )


async def scan_pipeline_image(
    image_bytes: bytes,
    context: PipelineAIContext,
    pig_run_id: str,
    chainage_m: float | None,
    operator_id: str,
    pipeline_segment_id: str,
    is_hca: bool,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a pipeline integrity AI image for adversarial content.

    Fail-safe contract: AdversarialPipelineImageError or httpx error →
    suppress AI defect classification, route to qualified ILI analyst for
    manual review per API 1160 Section 6.2.2. Do not issue FFS assessment
    or generate repair priority without clean scan for HCA segments.

    Args:
        image_bytes: MFL heatmap render, UT B-scan image, drone RGB frame,
            or ILI camera pig frame bytes.
        context: PipelineAIContext identifying the inspection modality.
        pig_run_id: ILI run identifier from the inspection vendor.
        chainage_m: Axial pipe position in metres from reference launcher.
        operator_id: PHMSA operator ID.
        pipeline_segment_id: Operator segment identifier.
        is_hca: True if this chainage is within a PHMSA-defined HCA.
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict.

    Raises:
        AdversarialPipelineImageError: 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"pipeline:{context.value}:{pig_run_id}:{chainage_m}",
        "metadata": {
            "pig_run_id": pig_run_id,
            "chainage_m": chainage_m,
            "operator_id": operator_id,
            "pipeline_segment_id": pipeline_segment_id,
            "is_hca": is_hca,
            "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_pipeline_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        pig_run_id=pig_run_id,
        chainage_m=chainage_m,
        is_hca=is_hca,
        flagged=result["score"] > PIPELINE_AI_THRESHOLD,
    )

    if result["score"] > PIPELINE_AI_THRESHOLD:
        raise AdversarialPipelineImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            pig_run_id=pig_run_id,
            chainage_m=chainage_m,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def _write_pipeline_scan_audit(
    *, image_hash: str, scan_id: str, score: int,
    context: PipelineAIContext, pig_run_id: str,
    chainage_m: float | None, is_hca: bool, flagged: bool,
) -> None:
    record = {
        "ts": datetime.now(timezone.utc).isoformat(),
        "scan_id": scan_id,
        "image_sha256": image_hash,
        "context": context.value,
        "score": score,
        "threshold": PIPELINE_AI_THRESHOLD,
        "flagged": flagged,
        "pig_run_id": pig_run_id,
        "chainage_m": chainage_m,
        "is_hca": is_hca,
        "regulatory_refs": [
            "PHMSA 49 CFR 195.452 (Hazardous Liquid Integrity Management)",
            "PHMSA 49 CFR 192.917 (Natural Gas Integrity Management)",
            "API 1160 (Managing System Integrity for Hazardous Liquid Pipelines, 3rd ed.)",
            "API RP 1176 (Assessment and Management of Cracking in Pipelines, 1st ed.)",
            "ASME B31.8S (Managing System Integrity of Gas Pipelines)",
            "NACE SP0169-2013 (Control of External Corrosion on Underground Metallic Piping)",
            "PHMSA 49 CFR 192.613 (Emergency excavation response requirements)",
        ],
    }
    audit_path = Path("/var/log/glyphward/pipeline_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_pipeline_image at each ILI AI rendered-image ingestion boundary: before MFL heatmap AI (threshold 40), before UT B-scan AI (threshold 40), before drone corrosion RGB frame AI (threshold 40), and before ILI camera pig frame AI (threshold 40). On AdversarialPipelineImageError: suppress AI defect classification, route the image and chainage position to a qualified ILI data analyst for manual review per API 1160 Section 6.2.2. For HCA segments: do not generate fitness-for-service assessments or repair priority scores based on adversarially flagged images without human review. Get early access

Related questions

Further reading

• Prompt injection in transmission line inspection AI — EPRI SmartLine, drone RGB+thermal+UV inspection AI adversarial attacks
• Prompt injection in smart manufacturing ICS AI — SCADA process camera AI and quality inspection adversarial attacks
• Prompt injection in offshore platform inspection AI — ROV camera corrosion AI and riser inspection adversarial attacks
• Prompt injection in dam safety monitoring AI — seepage detection AI and spillway inspection adversarial attacks
• Prompt injection scanning API free tier — 10 scans/day, no card required