Prompt injection in subsea and offshore inspection AI

Four adversarial injection surfaces in subsea and offshore inspection AI pipelines

1. ROV pipeline inspection AI bypass (Oceaneering AI, TechnipFMC AI, Subsea7 inspection AI)

ROV-based subsea pipeline inspection generates continuous video footage of pipeline surfaces at depths ranging from 50 metres in shallow offshore fields to over 3,000 metres in ultra-deepwater developments. Oceaneering’s AI-assisted pipeline inspection system processes ROV camera footage at 25–30 frames per second, submitting sequential frame images to convolutional neural network classifiers trained to detect corrosion pitting, external corrosion patches, mechanical damage from anchor drag or trawl gear impact, marine growth (sponges, coral, barnacles) masking underlying pipe surface condition, and free-span conditions where the pipeline is unsupported across a seabed depression. The AI classifier’s output — anomaly present/absent, anomaly type, severity classification — feeds into the inspection finding database used to prioritize repair and maintenance activities and to generate the regulatory compliance documentation that operators must submit to the Norwegian PSA, UK OPRED (Offshore Petroleum Regulator for Environment and Decommissioning), BSEE (Bureau of Safety and Environmental Enforcement) for US Gulf of Mexico operations, and equivalent regulators globally. Subsea7’s inspection, repair, and maintenance (IRM) vessel operations incorporate AI-assisted ROV footage analysis that generates defect reports used to plan contractor dive and intervention workscopes — $50,000–$200,000 mobilization cost per ROV intervention means AI-driven defect prioritization has direct operational budget implications.

The adversarial attack against ROV pipeline inspection AI targets the ROV video frame images at the topside data processing stage — the point at which video from the ROV umbilical is ingested into the vessel or shore-based analysis workstation before submission to the AI classifier. ROV video is transmitted from the subsea vehicle to the surface vessel through fiber optic or coaxial umbilical cable, converted to digital format at the topside termination unit, and streamed to the AI analysis workstation running the inspection software. An adversary with access to the topside data processing system — an insider on the inspection vessel, a compromised inspection software update, or a supply chain attack on the inspection analysis platform — can apply adversarial pixel perturbations to the ROV video frame stream before it reaches the AI classifier. Adversarial perturbations designed to suppress corrosion defect detection cause the AI to output “no anomaly” classifications for frames showing external corrosion that should trigger remediation priority elevation. For pipelines subject to corrosion-accelerated failure mechanisms — particularly in high-CO2 and H2S production fluid service where corrosion rates can reach 5–10 mm/year — adversarial suppression of early-stage corrosion findings delays remediation by one or more inspection cycle intervals (typically 1–5 years between ROV inspection campaigns), potentially allowing corrosion to progress to wall penetration and hydrocarbon release. The Deepwater Horizon blowout (2010, $65 billion in damages) and Pipelines de Pase substation fire demonstrate the magnitude of consequence when subsea integrity management failures lead to catastrophic release events.

BSEE regulations under 30 CFR Part 250 Subpart Q require operators of OCS (Outer Continental Shelf) pipelines to implement inspection programs that detect and report safety anomalies; inspection finding reports generated by AI analysis that have been adversarially manipulated do not satisfy the regulatory inspection documentation requirement because the inspection did not detect anomalies that were present in the inspected pipeline. BSEE civil penalties under 30 USC §1350(b) for violations include up to $40,000 per day per violation; criminal penalties under 30 USC §1350(c) for knowing and willful violations reach $100,000 per day plus imprisonment. Norwegian PSA (Petroleum Safety Authority) regulations under Framework Regulations §9 require operators to establish and maintain an integrated management system ensuring that facilities are operated safely; AI-based inspection systems that are susceptible to adversarial manipulation without detection mechanisms create a management system integrity gap cognizable under Framework Regulations §9 and Activities Regulations §47 (barrier management).

2. Sonar and acoustic imaging AI injection (Kongsberg Maritime AI, RESON sonar AI, Blueprint Subsea)

Multibeam sonar and side-scan sonar systems deployed on ROVs, AUVs, and inspection vessels generate acoustic imaging data that is rendered as visual grayscale or false-color images for AI analysis — the sonar scan visualization is a 2D image representation of the acoustic backscatter intensity pattern from the seabed surface or pipeline outer surface, processed identically to optical camera images by AI visual classification models. Kongsberg Maritime’s HISAS (Interferometric Synthetic Aperture Sonar) system, deployed on Kongsberg HUGIN AUVs, produces sonar imagery at centimeter resolution that AI models process to detect pipeline free-spans, seabed scour, anchor drag marks, and pipeline burial depth changes. RESON (now part of Teledyne) and Blueprint Subsea sonar systems generate similar sonar visualization images used in AI-assisted seabed and pipeline condition assessment. The sonar visualization image is functionally equivalent to an optical photograph for AI processing purposes: it is a 2D raster image containing spatial information about the imaged subject, processed by the same classes of convolutional neural networks used for optical image analysis, with the same adversarial vulnerability to pixel-layer perturbations.

The adversarial attack against sonar imaging AI is architecturally identical to the ROV camera injection attack: adversarial perturbations applied to the rendered sonar visualization image at the topside data processing stage can cause the AI to suppress detection of pipeline free-span conditions shown in the sonar image, misclassify seabed scour extent affecting pipeline burial depth, or suppress anchor drag mark detection indicating potential pipeline damage from third-party maritime activity. Free-span conditions in offshore pipelines are a primary cause of fatigue failure — unsupported spans vibrate under current loading, and vibration-induced fatigue accumulates toward a span length and current velocity threshold that causes pipe wall cracking and eventual fracture. DNV-RP-F105 (Free Spanning Pipelines) and ISO 13628-1 provide the industry guidance on free-span assessment that integrity engineers apply to AI-generated sonar inspection findings; adversarial suppression of free-span detection from sonar AI introduces free-span fatigue exposure that the DNV-RP-F105 analysis never receives the triggering input to assess. DNV (formerly DNV GL) acts as the primary offshore certification body for pipeline integrity management programs in Norway, the North Sea, and globally; certification against DNVGL-ST-F101 (Submarine Pipeline Systems) and equivalent standards requires inspection program adequacy that adversarially compromised AI analysis cannot provide.

3. FPSO hull and offshore platform corrosion AI injection (Baker Hughes BHGE AI, Aker BP AI, Equinor digital inspection AI)

Floating Production Storage and Offloading (FPSO) vessels are the primary production infrastructure for deep-water and marginally located offshore fields, with more than 200 FPSOs in operation globally holding oil and gas storage capacities up to 2.2 million barrels. FPSO hull integrity is safety-critical: hull plate corrosion and fatigue cracking below the waterline create catastrophic oil spill risk and structural failure risk; hull inspection requires ROV-based external surveys and periodic dry-dock internal inspection. Baker Hughes Digital Solutions operates an AI-assisted inspection platform that processes drone and ROV camera images from topside facility and FPSO hull inspection campaigns, applying computer vision AI to classify corrosion severity, coating breakdown, structural crack indicators, and mechanical damage in inspection images submitted by asset operators globally. Aker BP and Equinor have developed AI models for automated analysis of digital inspection images from Norwegian Continental Shelf platforms and FPSOs, generating structured finding databases that integrity management engineers use in risk-based inspection (RBI) programs following API RP 580 and API RP 581 risk-based inspection methodology. Offshore platform topside facility inspection — structural steel corrosion mapping, process piping external inspection, pressure vessel external visual inspection — has similarly adopted AI-assisted image analysis using drone-captured inspection photographs submitted to AI classifiers for corrosion severity rating according to ISO 8501-1 (preparation of steel substrates before application of paints and related products) and equivalent international corrosion rating standards.

The adversarial attack against FPSO hull and platform inspection AI targets the inspection photograph at the point of upload from the drone or ROV data downloader to the AI analysis platform. Inspection photographs from offshore drone surveys and ROV campaigns are downloaded from the camera storage, pre-processed (exposure correction, image stabilization, file format conversion), and uploaded to cloud-based AI analysis platforms. Adversarial perturbations applied at the pre-processing and upload stage — through a compromised pre-processing software component or a supply chain attack on the inspection software update — can cause the AI to classify active corrosion patches as light surface rust (downgrading severity from Grade C to Grade B under ISO 8501-1), miss paint film breakdown indicating imminent corrosion onset, or suppress structural crack indication detection in high-consequence areas. The financial consequence of adversarial FPSO hull inspection AI manipulation is compounded by the cost structure of FPSO inspection: dry-dock inspection costs $50–$200 million for a large FPSO and is conducted every 5–10 years; adversarial suppression of corrosion findings that would have required dry-dock intervention defers a $50–$200 million inspection and repair cost at the risk of allowing structural deterioration to progress to catastrophic failure. The BP oil spill liability framework (OPCRP, OPA 90, OSPAR) and the Equinor and BP internal liability reserves for major FPSO incidents demonstrate the magnitude of consequence that corrosion inspection AI failure can ultimately produce.

4. Subsea production system AI inspection fraud (TechnipFMC AI, SLB OneSubsea AI, Aker Solutions digital twin AI)

Subsea production systems — subsea trees, manifolds, flowline jumpers, pipeline end manifolds (PLEMs), and pipeline end terminals (PLETs) — are the critical well-control and production infrastructure on the seabed that must be inspected regularly to verify valve position status, connector integrity, hydraulic line routing, and structural condition. TechnipFMC’s digital inspection services apply AI to ROV camera footage from subsea tree inspection visits to classify subsea tree valve position indicators (open/closed status verification), subsea control module condition, and structural anomaly presence. SLB (Schlumberger) OneSubsea AI and Aker Solutions’ digital twin platforms process ROV inspection footage against 3D digital twin models of the subsea production system, using AI to compare the inspected as-is condition against the design as-built state and flag deviations that indicate component degradation, misalignment, or damage. Wood Group (now John Wood Group) and Applus+ operate AI-assisted inspection review services where ROV footage is submitted to cloud platforms for automated analysis generating inspection finding reports used by operators for regulatory compliance submissions.

The adversarial attack against subsea production system inspection AI creates a compliance fraud vector that is commercially motivated by the high cost of subsea intervention. Subsea intervention — deploying a light work ROV or heavy work ROV vessel to perform a subsea corrective maintenance operation — costs $500,000–$3 million per vessel day depending on water depth and vessel specification. When AI inspection findings identify a subsea tree hydraulic leak, a connector seal degradation, or a structural anomaly requiring intervention, the operator faces immediate intervention mobilization cost. Adversarial manipulation of the ROV inspection footage submitted to the AI before it reaches the TechnipFMC or SLB AI analysis platform — suppressing the anomaly detection that would trigger an intervention work order — defers the intervention cost at the risk of allowing the underlying condition to deteriorate. For small operators managing marginal fields with limited capital, this cost deferral motivation creates a plausible adversarial manipulation incentive that is operationally distinct from the external attacker motivation applicable to most industrial AI adversarial threat models. Norwegian PSA Activities Regulations §86 (inspection and testing) and BSEE 30 CFR Part 250.801 (Well Control Equipment Inspection) impose mandatory inspection interval and documentation requirements that adversarially manipulated AI inspection reports cannot satisfy, creating regulatory fraud exposure for operators who submit adversarially generated AI inspection findings to meet regulatory inspection documentation obligations.

Integration: subsea inspection AI image ingestion with Glyphward pre-scan

The Glyphward scan gate belongs at the inspection image ingestion point in each subsea AI pipeline — before the ROV camera frame, sonar visualization, platform inspection photograph, or subsea production system image is passed to the AI analysis engine. The async pattern below handles all four subsea inspection contexts through a shared scan_subsea_ai_image function, with thresholds calibrated to the safety-critical nature of subsea integrity findings and structured audit output suitable for BSEE inspection compliance documentation and Norwegian PSA/UK HSE integrity management audit records.

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"

# Per-context thresholds for subsea inspection AI
ROV_PIPELINE_INSPECTION_THRESHOLD   = 55  # Oceaneering / TechnipFMC / Subsea7 ROV AI
SONAR_ANOMALY_DETECTION_THRESHOLD   = 55  # Kongsberg / RESON / Blueprint Subsea sonar AI
FPSO_PLATFORM_INSPECTION_THRESHOLD  = 55  # Baker Hughes / Aker BP / Equinor inspection AI
SUBSEA_PRODUCTION_COMPLIANCE_THRESHOLD = 60  # TechnipFMC / SLB OneSubsea compliance AI


class SubseaInspectionAIContext(Enum):
    ROV_PIPELINE_INSPECTION   = "rov_pipeline_inspection"    # threshold 55
    SONAR_ANOMALY_DETECTION   = "sonar_anomaly_detection"    # threshold 55
    FPSO_PLATFORM_INSPECTION  = "fpso_platform_inspection"   # threshold 55
    SUBSEA_PRODUCTION_COMPLIANCE = "subsea_production_compliance"  # threshold 60


_CONTEXT_THRESHOLDS: dict[SubseaInspectionAIContext, int] = {
    SubseaInspectionAIContext.ROV_PIPELINE_INSPECTION:     ROV_PIPELINE_INSPECTION_THRESHOLD,
    SubseaInspectionAIContext.SONAR_ANOMALY_DETECTION:     SONAR_ANOMALY_DETECTION_THRESHOLD,
    SubseaInspectionAIContext.FPSO_PLATFORM_INSPECTION:    FPSO_PLATFORM_INSPECTION_THRESHOLD,
    SubseaInspectionAIContext.SUBSEA_PRODUCTION_COMPLIANCE: SUBSEA_PRODUCTION_COMPLIANCE_THRESHOLD,
}


class AdversarialSubseaAIImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in a
    subsea inspection AI input image above the context threshold.

    Attributes:
        scan_id: Glyphward scan identifier for the audit record.
        score: Adversarial signal score (0-100).
        context: The SubseaInspectionAIContext in which detection occurred.
        flagged_region: Optional dict describing the flagged pixel region.
    """

    def __init__(
        self,
        scan_id: str,
        score: int,
        context: SubseaInspectionAIContext,
        flagged_region: dict | None = None,
    ) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial subsea inspection AI image detected: "
            f"context={context.value} score={score} scan_id={scan_id}"
        )


async def scan_subsea_ai_image(
    image_path: Path,
    context: SubseaInspectionAIContext,
    asset_id: str,
    inspection_campaign_id: str,
    operator_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a subsea inspection AI input image for adversarial pixel content.

    Args:
        image_path: Absolute path to the ROV frame / sonar / inspection image.
        context: SubseaInspectionAIContext enum value identifying the AI pipeline.
        asset_id: Pipeline or facility asset identifier (e.g. pipeline segment ID).
        inspection_campaign_id: Inspection campaign reference for audit correlation.
        operator_id: Asset operator identifier for regulatory compliance audit trail.
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict: scan_id, score, flagged_region, modality.

    Raises:
        AdversarialSubseaAIImageError: if score exceeds context threshold.
        httpx.HTTPStatusError: on Glyphward API errors (fail-closed: quarantine image).
    """
    threshold = _CONTEXT_THRESHOLDS[context]
    image_bytes = image_path.read_bytes()
    image_hash = hashlib.sha256(image_bytes).hexdigest()

    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"subsea:{context.value}:{inspection_campaign_id}",
        "metadata": {
            "asset_id": asset_id,
            "inspection_campaign_id": inspection_campaign_id,
            "operator_id": operator_id,
            "image_sha256": image_hash,
        },
    }

    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=5.0,
    )
    resp.raise_for_status()
    result = resp.json()

    await write_subsea_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        threshold=threshold,
        asset_id=asset_id,
        inspection_campaign_id=inspection_campaign_id,
        operator_id=operator_id,
        flagged=result["score"] > threshold,
    )

    if result["score"] > threshold:
        raise AdversarialSubseaAIImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            flagged_region=result.get("flagged_region"),
        )

    return result


async def write_subsea_scan_audit(
    *,
    image_hash: str,
    scan_id: str,
    score: int,
    context: SubseaInspectionAIContext,
    threshold: int,
    asset_id: str,
    inspection_campaign_id: str,
    operator_id: str,
    flagged: bool,
) -> None:
    """Append structured JSON audit record to subsea inspection AI scan log.

    Satisfies BSEE 30 CFR Part 250 inspection documentation requirements,
    Norwegian PSA Activities Regulations §86 integrity management documentation,
    and UK HSE PFEER regulation inspection program audit evidence needs.
    """
    record = {
        "ts": datetime.now(timezone.utc).isoformat(),
        "scan_id": scan_id,
        "image_sha256": image_hash,
        "context": context.value,
        "score": score,
        "threshold": threshold,
        "flagged": flagged,
        "asset_id": asset_id,
        "inspection_campaign_id": inspection_campaign_id,
        "operator_id": operator_id,
    }
    audit_path = Path("/var/log/glyphward/subsea_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")


async def process_subsea_image_batch(
    images: list[tuple[Path, SubseaInspectionAIContext, str, str, str]],
) -> list[dict]:
    """Process a batch of (path, context, asset_id, campaign_id, operator_id) tuples."""
    async with httpx.AsyncClient() as client:
        tasks = [
            scan_subsea_ai_image(
                image_path=path,
                context=ctx,
                asset_id=aid,
                inspection_campaign_id=cid,
                operator_id=oid,
                client=client,
            )
            for path, ctx, aid, cid, oid in images
        ]
        results = []
        for coro in asyncio.as_completed(tasks):
            try:
                results.append(await coro)
            except AdversarialSubseaAIImageError as exc:
                results.append({
                    "status": "quarantined",
                    "context": exc.context.value,
                    "scan_id": exc.scan_id,
                    "score": exc.score,
                    "flagged_region": exc.flagged_region,
                })
        return results

Deploy scan_subsea_ai_image at the image ingestion boundary: before ROV video frames reach Oceaneering or TechnipFMC pipeline inspection AI; before sonar visualization images reach Kongsberg HISAS or RESON sonar AI; before platform inspection photographs reach Baker Hughes or Equinor FPSO inspection AI; and before subsea tree inspection images reach SLB OneSubsea or TechnipFMC compliance AI. The audit log’s asset_id and inspection_campaign_id fields create the evidence chain linking adversarial scan records to specific inspection campaigns for BSEE and Norwegian PSA inspection documentation compliance. Get early access

Coverage matrix

Related questions

Further reading

• Prompt injection in energy and utilities field operations AI — pipeline inspection and grid maintenance
• Prompt injection in renewable energy AI — wind blade and solar thermal inspection
• Prompt injection in drone and UAV inspection AI — infrastructure inspection and survey
• Prompt injection in maritime shipping AI — vessel inspection and cargo monitoring
• Prompt injection scanning API free tier — 10 scans/day, no card required