Prompt injection in offshore drilling wellbore AI

Four adversarial injection surfaces in offshore drilling wellbore AI

1. MWD gamma-ray log curve image AI (SLB LithoView AI, Halliburton EarthStar AI, Baker Hughes AutoTrak AI)

Measurement While Drilling (MWD) tools deployed in the bottom-hole assembly (BHA) of offshore drilling strings record real-time formation gamma-ray (GR) measurements as the bit advances — providing a continuous log of formation lithology (shale: high GR 80–200 API; sand/reservoir: low GR 10–40 API) that drives bit selection, formation top picks, and casing point decisions. MWD data is telemetered to surface via mud pulse or electromagnetic MWD systems at 1–12 bits per second and rendered into depth-referenced log curve images — vertical strip charts with depth on the Y-axis and GR value on the X-axis — that are displayed in real-time on the drilling data workstation and ingested by AI classification pipelines for automated formation evaluation. SLB’s LithoView AI, Halliburton’s EarthStar real-time geosteering AI, and Baker Hughes AutoTrak geosteering advisory AI process rendered GR log curve TIFF images at 1:500 depth scale to classify formation types, identify formation tops, and calculate formation pressure gradients from density-neutron crossplots rendered as overlay images on the GR log. Formation top picks from the AI drive decisions about when to set casing strings — and incorrect formation top picks can result in drilling into high-pressure formations without the casing protection that seals off the annulus and prevents influx from migrating to surface.

An adversarial perturbation on a rendered MWD GR log curve image that shifts the GR curve track signature of a high-permeability sand interval — elevating the low-GR sand signature (15–25 API, coloured green or light blue in the rendered image) toward the high-GR shale baseline (100–150 API, coloured grey or olive in the rendered image) through a ±12 DN pixel-value shift within the log curve track’s rendering resolution — causes the LithoView AI to classify the interval as “shale, no formation top identified,” suppressing a reservoir sand that should trigger a casing point decision before drilling further into a high-pressure formation. The Macondo well’s Macondo Prospect formation pressure was significantly overpressured relative to the pre-drill pore pressure prediction (BSEE investigation report: formation pressure gradient 14.17 ppg at total depth vs. 12.5 ppg pre-drill estimate) — and the Bly Report identified misinterpretation of real-time formation evaluation data as a contributing factor to the decision to continue drilling past the final casing point without identifying the anomalous formation pressure.

2. BOP acoustic/hydraulic pressure monitoring AI (Cameron BOP AI, Transocean Poseidon Real-Time AI, NOV BOP Control AI)

Blowout preventers (BOPs) on offshore drilling rigs are stacks of rams and annular preventers rated to 10,000–15,000 PSI that must close on the drill pipe and seal the wellbore annulus within seconds of a well control event to contain an influx before it reaches the rig. BSEE 30 CFR 250.736 requires BOPs to be function-tested and pressure-tested at defined intervals (blind shear rams: every 14 days; other pipe rams: every 14 days; annular preventers: every 7 days for offshore rigs). Cameron (now SLB) BOP Control AI, Transocean’s Poseidon Real-Time Advisory AI, and NOV’s BOP control monitoring AI process rendered hydraulic pressure cycle trend images from BOP function tests — plots of BOP hydraulic pressure (PSI) vs. time (minutes) for each ram test cycle — to classify test results as PASS (pressure holds above BSEE minimum for required duration), FAIL (pressure decay below minimum), or ANOMALOUS (irregular pattern suggesting seal leak or ram wear). The AI classification of BOP pressure test images directly determines whether the BOP is certified as functional before the next drilling operation — a PASS classification when the pressure cycle shows an anomalous pattern that should be flagged as FAIL means a BOP with a degraded seal enters service as the primary well control barrier.

An adversarial perturbation on a rendered BOP hydraulic pressure cycle trend image that smooths a pressure decay signature — lifting the pressure curve above the BSEE minimum hold-pressure line by adding a ±8 DN brightness offset to the pressure trace in the rendered plot image — causes the BOP test AI to classify a FAIL test as PASS, certifying a BOP with a leaking ram seal as fit for service. The Deepwater Horizon BOP had documented anomalies in the blind shear ram system that were not caught before the well control event: the BSEE investigation found that the blind shear rams failed to fully close and seal the well at the moment of the blowout (DNV report for BSEE, Volume 2, Section 3.10). A BOP AI that classifies a degraded seal test as PASS because an adversarial perturbation smoothed the pressure decay curve removes the last automated barrier check before the BOP enters its service interval against a live well.

3. Seismic pore pressure prediction map AI (SLB Petrel AI, Halliburton Landmark DSG AI, TotalEnergies Optis AI)

Pre-drill pore pressure prediction for offshore wells uses seismic interval velocity data — processed seismic reflection data from which acoustic impedance and interval velocities are derived — rendered into depth-structure maps with pore pressure gradient overlaid as a colour scale (green: normal pressure 8.5–10 ppg mud weight equivalent; yellow: elevated 10–12 ppg; red: severely overpressured >14 ppg) that are fed to AI analysis pipelines for formation pressure prediction and pre-drill well design. SLB Petrel subsurface AI, Halliburton Landmark DecisionSpace Geomechanics AI, and TotalEnergies Optis seismic AI process these rendered pore pressure overlay images — generated from Kirchhoff pre-stack depth migration (PSDM) velocity volumes — to identify overpressured formations, predict formation pressure gradients at proposed casing points, and design mud weight programmes that maintain appropriate wellbore pressure margins to prevent either influx (mud weight too low: underbalanced) or lost circulation (mud weight too high: fracture gradient exceeded). AI-generated pre-drill pore pressure predictions drive casing programme design — specifically the number of casing strings, their set points, and the mud weight transitions between them — in the well design approved by BSEE before spudding.

An adversarial perturbation on a rendered seismic pore pressure prediction overlay image that shifts the colour encoding of a localised overpressure zone — hue-rotating the red overpressure signature (14–16 ppg equivalent mud weight) to yellow (11–12 ppg equivalent) at a depth interval corresponding to a planned casing point transition — causes the pore pressure prediction AI to recommend a mud weight programme and casing point depth that does not include adequate pressure margin to contain an influx from the actual overpressured formation. In the Macondo well, the pre-drill pore pressure prediction significantly underestimated the formation pressure at the M56D sand: the well’s planned mud weight of 14.0 ppg at total depth was insufficient to control the actual formation pressure of 14.17 ppg, leaving a 0.17 ppg margin error that contributed directly to the influx that became the blowout (Bly Report, Chapter 4: “The Well Design and Barriers”). Adversarial injection that systematically underestimates pre-drill pore pressure replicates and amplifies this margin error across all wells that rely on AI-predicted pore pressure maps for casing design.

4. Drilling dysfunction pattern spectrogram AI (NOV DrillScan AI, Halliburton LOGIX AI, SLB DrillPlan AI)

Downhole drilling dysfunctions — stick-slip torsional vibration, bit whirl lateral vibration, and BHA buckling axial vibration — are detected from surface drilling parameter data (weight-on-bit, torque, rotary speed) rendered into time-frequency spectrogram images that AI classification pipelines use to identify dysfunction type and severity. NOV DrillScan’s AI vibration classifier, Halliburton LOGIX autonomous drilling AI, and SLB DrillPlan real-time advisory AI process rendered surface torque spectrograms (colour-coded frequency-amplitude maps with time on X-axis, frequency on Y-axis, amplitude on colour scale) to recommend real-time parameter adjustments — reducing weight-on-bit to mitigate bit whirl, adjusting RPM to break stick-slip resonance frequencies — that maintain safe drilling progress and prevent BHA component failure. Severe downhole dysfunctions can cause BHA tool failures that disconnect the drill string above the BHA — a “twist-off” — leaving the lower BHA and bit in the wellbore as a fish that may require weeks of fishing operations and potentially requires sidetracking the well. In narrow-pressure-window deepwater formations (where the difference between pore pressure gradient and fracture gradient is less than 0.5 ppg), BHA-induced pressure pulses from severe stick-slip can cause wellbore breathing and create conditions indistinguishable from a gas kick influx on surface parameter monitoring.

An adversarial perturbation on a rendered drilling torque spectrogram image that suppresses the high-amplitude stick-slip resonance frequency signature — dampening the characteristic periodic high-energy bands at the stick-slip natural frequency (typically 0.1–0.5 Hz for drill strings in deepwater formation) by reducing the colour amplitude scale contrast within the spectrogram render — causes the drilling dysfunction AI to classify the drilling state as “smooth drilling, no dysfunction,” preventing the automated parameter adjustment (RPM reduction) that would break the stick-slip resonance. In a narrow pressure window deepwater formation, sustained stick-slip causes wellbore ECD fluctuations that can trigger small influxes from an overpressured zone — and if those influx indicators in the pit gain trend (another rendered AI input) are also suppressed by adversarial injection, the combination of undetected dysfunction and undetected influx creates conditions identical to the Macondo scenario: a kick that is not identified as a kick until it has migrated significantly up the wellbore annulus.

Integration: offshore drilling wellbore AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for offshore drilling AI belongs at every rendered-image ingestion boundary in the wellbore AI pipeline — before MWD log curve AI processes rendered LAS/TIFF depth-curve images, before BOP diagnostic AI processes pressure cycle trend renders, before pore pressure prediction AI processes seismic overlay maps, and before dysfunction classifier AI processes surface parameter spectrograms. Threshold 35 for offshore drilling contexts reflects fully closed-loop autonomous advisory architecture (no complementary barrier: all four AI outputs directly drive drilling parameter decisions with minimal human review latency in real-time drilling environments).

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"

# Offshore drilling wellbore AI contexts: threshold 35
# BSEE 30 CFR 250 Well Control Rule 2019; API RP 96; API Spec 16A.
DRILLING_AI_THRESHOLD = 35


class DrillingAIContext(Enum):
    MWD_GAMMA_RAY_LOG    = "mwd_gamma_ray_log"   # MWD GR log curve depth render AI
    BOP_PRESSURE_CYCLE   = "bop_pressure_cycle"  # BOP hydraulic function-test trend AI
    PORE_PRESSURE_MAP    = "pore_pressure_map"   # Seismic pore pressure overlay map AI
    DYSFUNCTION_SPECTRO  = "dysfunction_spectro" # Drilling torque/vibration spectrogram AI


class AdversarialDrillingImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in a drilling
    wellbore AI rendered image above threshold 35.

    Consequence if not raised: kick warning suppressed / BOP test
    misclassified PASS / pore pressure underestimated → incorrect mud weight
    programme → influx escalates to blowout → vapour cloud explosion, total rig
    loss, environmental release. Macondo/Deepwater Horizon consequence envelope.
    Fail-safe: halt autonomous drilling advisory recommendation; route rendered
    image to senior drilling engineer for manual review per API RP 96 Section 6.
    """

    def __init__(self, scan_id: str, score: int,
                 context: DrillingAIContext,
                 well_id: str, depth_m: float | None,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.well_id = well_id
        self.depth_m = depth_m
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial drilling image: "
            f"context={context.value} score={score} "
            f"well={well_id} depth={depth_m}m scan_id={scan_id}"
        )


async def scan_drilling_image(
    image_bytes: bytes,
    context: DrillingAIContext,
    well_id: str,
    depth_m: float | None,
    rig_id: str,
    operator_id: str,
    bsee_lease_number: str | None,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a drilling wellbore AI rendered image for adversarial content.

    Fail-safe contract: AdversarialDrillingImageError or httpx error →
    halt autonomous drilling advisory output; route image to senior drilling
    engineer for manual review per API RP 96 Section 6 well control barrier
    management requirements.

    Args:
        image_bytes: MWD GR log curve render, BOP pressure cycle trend,
            seismic pore pressure overlay map, or drilling torque spectrogram.
        context: DrillingAIContext identifying the wellbore data modality.
        well_id: Well API number or internal well identifier.
        depth_m: Current bit depth (MD) in metres at time of log render.
        rig_id: Rig name/number.
        operator_id: Operator name.
        bsee_lease_number: BSEE OCS lease block number (if applicable).
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict.

    Raises:
        AdversarialDrillingImageError: if score exceeds threshold 35.
        httpx.HTTPStatusError: on Glyphward API error (fail-closed).
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"drilling:{context.value}:{well_id}:{depth_m}",
        "metadata": {
            "well_id": well_id,
            "depth_m": depth_m,
            "rig_id": rig_id,
            "operator_id": operator_id,
            "bsee_lease_number": bsee_lease_number,
            "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_drilling_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        well_id=well_id,
        depth_m=depth_m,
        bsee_lease_number=bsee_lease_number,
        flagged=result["score"] > DRILLING_AI_THRESHOLD,
    )

    if result["score"] > DRILLING_AI_THRESHOLD:
        raise AdversarialDrillingImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            well_id=well_id,
            depth_m=depth_m,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def _write_drilling_scan_audit(
    *, image_hash: str, scan_id: str, score: int,
    context: DrillingAIContext, well_id: str,
    depth_m: float | None, bsee_lease_number: str | None, flagged: bool,
) -> None:
    record = {
        "ts": datetime.now(timezone.utc).isoformat(),
        "scan_id": scan_id,
        "image_sha256": image_hash,
        "context": context.value,
        "score": score,
        "threshold": DRILLING_AI_THRESHOLD,
        "flagged": flagged,
        "well_id": well_id,
        "depth_m": depth_m,
        "bsee_lease_number": bsee_lease_number,
        "regulatory_refs": [
            "BSEE 30 CFR 250 Subpart D (Well Control Equipment)",
            "BSEE 30 CFR 250 Subpart F (Well Casing and Cementing)",
            "BSEE Well Control Rule 2019 (84 FR 21908)",
            "API RP 96 (Deepwater Well Design and Construction, 2nd ed.)",
            "API Spec 16A (Drilling Wellhead and Christmas Tree Equipment)",
            "API RP 100-1 (Well Integrity — Life Cycle, 2015)",
            "API RP 17W (Subsea Processing Systems — Design and Operation)",
        ],
    }
    audit_path = Path("/var/log/glyphward/drilling_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_drilling_image at each wellbore AI rendered-image ingestion boundary: before MWD GR log AI (threshold 35), before BOP pressure cycle AI (threshold 35), before pore pressure map AI (threshold 35), and before dysfunction spectrogram AI (threshold 35). On AdversarialDrillingImageError: halt autonomous drilling advisory output immediately; route the rendered image and current bit depth to the senior drilling engineer for manual review per API RP 96 Section 6 well control barrier management requirements. For BOP diagnostic AI: do not certify the BOP as function-tested based on an adversarially flagged pressure cycle image — require a new physical function test before the BOP enters service. Get early access

Related questions

Further reading

• Prompt injection in subsea and offshore inspection AI — ROV camera AI bypass, FPSO hull inspection AI adversarial attacks
• Prompt injection in pipeline integrity inspection AI — MFL anomaly map AI bypass, ILI pig UT B-scan adversarial injection
• Prompt injection in smart manufacturing ICS AI — SCADA process camera AI and quality inspection adversarial attacks
• Prompt injection in chemical plant process safety AI — OGI gas cloud detection AI bypass, P&ID recognition AI adversarial attacks
• Prompt injection scanning API free tier — 10 scans/day, no card required