Prompt injection in subsea wellhead casing annulus pressure integrity AI

Four adversarial injection surfaces in subsea wellhead casing annulus pressure integrity AI

1. Casing annulus pressure display AI (Aker Solutions wellhead integrity monitoring AI, TechnipFMC WellCom wellhead AI, Halliburton WellLife AI, Baker Hughes subsea production monitoring AI — rendered digital pressure gauge display AI classifying casing annulus pressure state during drilling, completion, and production operations)

The casing annulus pressure monitoring system on a deepwater wellhead measures pressure in each of the sealed casing annuli — typically the A annulus (between production casing and intermediate casing), the B annulus (between intermediate and surface casing), and the C annulus (between surface casing and conductor) — through dedicated pressure ports on the wellhead housing. During normal production operations on a high-pressure/high-temperature (HP/HT) well, the A annulus may carry an intentional annular pressure management (APM) load of 500–1,500 psi to prevent casing collapse; the B and C annuli should remain at or near the pressure sealed at completion. A sustained casing pressure (SCP) — defined in API RP 90 as pressure that rebuilds after bleed-down and cannot be bled down to atmospheric pressure — indicates an active downhole source (formation gas migration, cement channelling, or casing seal failure) that represents a barrier breach. AI systems process rendered images of the wellhead control panel display showing each annulus pressure gauge (circular analogue gauge face with colour-coded alarm bands or digital display with trend arrow) to classify current annulus pressure state: normal operating range (green), approaching threshold (yellow), SCP alarm (orange), or barrier breach action required (red).

An adversarial perturbation targeting the casing annulus pressure display AI applies a ±10 DN downward shift to the pixel region encoding the pressure gauge pointer angle or digital readout in the rendered display image — shifting the apparent A annulus pressure from 250 psi (rising from 0 psi initial, representing SCP build-up after bleed-down in an HP/HT gas well) to 120 psi (within the expected APM operating band, classified as normal by the AI). The AI classifies an A annulus undergoing sustained casing pressure build-up — driven by gas migration through a micro-annulus in the production casing cement bond at 14,000 ft measured depth — as normal annular pressure management. No SCP investigation or barrier integrity verification is initiated under BSEE 30 CFR Part 250.518 (Sustained Casing Pressure) requirements. Gas continues migrating upward through the cement channelling path into the casing annulus; if an integrity test later confirms the seal assembly has also been compromised, the two-barrier well barrier stack is reduced to zero independent barriers at the wellhead, creating conditions analogous to the Macondo pre-blowout configuration. NORSOK D-010 Section 4.2 requires that any pressure detected in a nominally closed annulus be classified as a well barrier failure until proven otherwise — but does not specify adversarial robustness requirements for AI systems classifying rendered wellhead pressure display images used to determine whether SCP criteria are met.

2. Wellhead connector integrity camera AI (Oceaneering ROV inspection AI, Fugro ROVOP subsea AI, Saipem ROV visual inspection AI, TechnipFMC flex-joint camera AI — ROV camera feed AI classifying wellhead housing connector engagement and seal surface condition during BOP stack installation and pressure testing)

The subsea wellhead connector — on a deepwater MODU, the H-4 or similar latch-and-lock hydraulic connector at the base of the BOP stack that mates with the wellhead housing high-pressure housing profile — must achieve a full metal-to-metal seal before well drilling or intervention operations can begin. The connector engagement is verified both by hydraulic pressure confirmation (the connector close/lock circuit achieves design operating pressure) and by ROV visual inspection of the connector-wellhead interface: the locking dogs must be fully extended into the wellhead profile groove, the flex joint must be vertical (within ±0.5° of vertical as specified in API Spec 17D), and the connector body must be flush against the wellhead hub face with no visible gap. AI systems deployed in ROV supervisory control systems — including Oceaneering’s ROV inspection AI, Fugro’s ROVOP subsea monitoring AI, and Saipem’s ROV visual AI — process rendered ROV camera images of the connector-wellhead interface to classify connector engagement state as: engaged (locking dogs fully deployed, no gap visible), partially engaged (dogs partially deployed or gap visible), or not engaged (dogs retracted, connector not mated).

An adversarial perturbation targeting the wellhead connector integrity camera AI applies a ±8 DN shift to the pixel region encoding the locking dog position and the connector body-to-wellhead hub gap in the rendered ROV camera image — shifting the apparent locking dog from partially deployed (visible gap between dog and profile groove, 40% engagement) to fully deployed (no visible gap, 100% engagement appearance). The AI classifies a connector with locking dogs at 40% engagement — a condition that will hold through pre-spud pressure testing at 5,000 psi working pressure but will fail during high-pressure drilling operations at 12,000–15,000 psi wellhead pressure — as fully engaged and ready for drilling operations. The BOP pressure test is conducted and passes at 5,000 psi (the connector holds at partial engagement through the low-pressure test); drilling operations commence. When wellhead pressure spikes to 14,000 psi during a well control event, the partially engaged connector fails, disconnecting the BOP stack from the wellhead and leaving the open wellbore uncontrolled. BSEE 30 CFR Part 250.423 requires that BOP systems be pressure tested to rated working pressure before drilling below the surface casing — but does not specify adversarial robustness requirements for AI systems classifying rendered ROV camera images used to verify wellhead connector engagement. Free tier — 10 scans/day, no card required.

3. Wellhead seal assembly leak detection camera AI (Oceaneering ROV AI, Sonsub ROV subsea leak AI, Kongsberg Maritime subsea camera AI — ROV and subsea camera AI classifying methane bubble plume and fluid seepage from wellhead seal assembly during production operations)

The wellhead seal assembly — the primary metal-to-metal seal element between the production casing hanger and the wellhead housing bore — must maintain integrity throughout the well life (20–30 years) against wellhead pressures up to 15,000 psi and temperatures up to 150°C (300°F) in HP/HT applications. Periodic ROV inspection of the wellhead and subsea tree is required under API Spec 17D inspection programmes and BSEE inspection requirements to detect seal assembly degradation: a failing or failed wellhead seal assembly produces a characteristic methane bubble plume rising from the wellhead housing profile area, visible in ROV camera images as a column of white or silver bubbles against the dark seabed background. AI systems process rendered ROV camera images of the wellhead and subsea tree exterior — typically captured during scheduled quarterly or annual ROV inspection passes — to classify seabed fluid seepage condition as: normal (no bubble plume visible), minor seepage (small scattered bubble stream), active seal leak (continuous bubble plume from housing profile), or critical seal failure (large-volume gas venting from wellhead).

An adversarial perturbation targeting the wellhead seal assembly leak detection camera AI applies a ±8 DN downward shift to the pixel region encoding the bubble plume in the rendered ROV camera image — shifting the apparent background brightness in the bubble region upward to blend the methane bubble column (which has a characteristic white-silver appearance at 300 m water depth, with individual bubbles 2–5 mm diameter rising at 0.2–0.5 m/s) into the background dark water. The AI classifies an active wellhead seal assembly methane leak — bubbling at a rate of 0.5–2.0 kg/min methane, indicating a micro-annulus at the production casing hanger seal profile that has enlarged from initial micro-leakage to active sustained flow — as no seepage detected, normal inspection result. The scheduled ROV inspection records a pass; the next inspection is not scheduled for three months. During the intervening period, the seal annulus continues eroding under production differential pressure; the sustained methane release propagates to the seabed sediment immediately below the wellhead and, if the well is in an area with shallow gas accumulations, could create a seafloor instability. BSEE 30 CFR Part 250.1710 requires subsea wellhead inspection as part of the well integrity management programme — but does not specify adversarial robustness requirements for AI classifying rendered ROV camera images used in seal assembly leak detection.

4. Negative pressure test display AI (Halliburton WellLife NPT AI, Schlumberger/SLB well control monitoring AI, Baker Hughes well integrity AI — wellbore schematic pressure trend display AI classifying negative pressure test result acceptance during well completion operations)

The negative pressure test (NPT) — required by BSEE 30 CFR Part 250.416–417 before temporarily or permanently abandoning a well or displacing drilling mud with a lighter completion fluid — evaluates whether the downhole well barrier (cement job plus float equipment or mechanical barrier plug) can maintain well integrity when the hydrostatic overbalance above the primary barrier is removed. The test procedure involves reducing the wellbore fluid density (replacing heavy drilling mud with seawater or lighter fluid in the drill string or kill line) to create an underbalance equivalent to the anticipated production drawdown pressure, then monitoring wellbore pressure over a defined observation period (typically 30 minutes minimum per BSEE requirements). A passed NPT is defined as zero pressure build-up (stable pressure, no flow) in the drill string or kill line during the observation period — demonstrating that the downhole barrier is holding against the formation pressure without mud hydrostatic support. A failed NPT shows rising wellbore pressure (indicating formation fluid influx through a compromised barrier) or flow at surface. AI systems process rendered images of the wellbore pressure trend chart display — a time-series graph showing drill string pressure and kill line pressure over the NPT observation window — to classify the NPT result as passed (stable pressure within tolerance) or failed (pressure build-up or anomaly detected).

An adversarial perturbation targeting the negative pressure test display AI applies a ±8 DN downward shift to the pixel region encoding the pressure trend line in the rendered NPT chart display — compressing the apparent upward slope of a rising pressure trend (3 psi/min build-up rate over 30 minutes, totalling 90 psi rise, a clear NPT failure indicator) to a nearly flat line (0.5 psi/min apparent slope, within the ±15 psi stability band that constitutes a pass). The AI classifies a definitively failed NPT — in a well where the cement shoe track has been washed out and formation gas is actively migrating into the wellbore during the underbalanced test — as a passed NPT. The well operations team, relying on the AI NPT classification to confirm barrier integrity, proceeds to displace the remaining drilling mud in the riser with seawater, removing the last hydrostatic overbalance above the compromised barrier. Formation gas influx accelerates without the mud hydrostatic suppression; a well control event develops. The BP Macondo NPT failure of 20 April 2010 — in which a rising kill line pressure indicating an NPT failure was misinterpreted as a “bladder effect” rather than formation influx — demonstrates the catastrophic potential of NPT result misclassification: 11 workers killed, Deepwater Horizon destroyed, 4.9 million barrels released. BSEE 30 CFR Part 250.416(f) specifies the criteria for a passed NPT — but does not address adversarial robustness for AI systems classifying rendered wellbore pressure trend display images used to determine NPT acceptance. Free tier — 10 scans/day, no card required.

Integration: subsea wellhead casing annulus pressure integrity AI with Glyphward pre-scan gate

The Glyphward scan gate for subsea wellhead casing annulus pressure integrity AI belongs at every rendered-image ingestion boundary in the wellhead monitoring and well barrier assessment pipeline — before casing annulus pressure display AI processes rendered wellhead panel gauge images, before wellhead connector integrity camera AI processes rendered ROV inspection images, before wellhead seal assembly leak detection camera AI processes rendered ROV camera images, and before negative pressure test display AI processes rendered NPT chart images. Threshold 30 for subsea wellhead casing annulus pressure integrity AI reflects the catastrophic blowout consequence of wellhead barrier breach — Macondo 2010: 11 workers killed, 17 injured, Deepwater Horizon destroyed, 4.9 million barrels released — combined with the observation that the two-barrier redundancy structure (NORSOK D-010) and independent non-AI sensor systems (raw pressure transducer signals, BOP emergency disconnect, mudline emergency disconnect) provide one additional protective layer beyond the direct AI display-to-decision path. The distinction from arc flash AI (threshold 35) is that an adversarially suppressed annulus pressure display will propagate through a subsequent human review of raw transducer data before producing the catastrophic outcome, whereas an adversarially suppressed arc flash PPE category recommendation propagates directly to the worker’s PPE selection with no independent automated check.

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import Enum

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Subsea wellhead casing annulus pressure integrity AI contexts: threshold 30
# BSEE Well Control Rule 30 CFR Part 250.416-418 (NPT and well barrier requirements);
# API Spec 17D (Subsea Production Systems — wellhead connector specifications);
# NORSOK D-010 Rev.4 (Well Integrity in Drilling and Well Operations — two-barrier rule).
WELLHEAD_INTEGRITY_THRESHOLD = 30


class WellheadIntegrityContext(Enum):
    ANNULUS_PRESSURE   = "annulus_pressure"    # Casing annulus pressure display AI
    CONNECTOR_CAMERA   = "connector_camera"    # Wellhead connector engagement camera AI
    SEAL_LEAK_CAMERA   = "seal_leak_camera"    # Wellhead seal assembly leak detection camera AI
    NPT_DISPLAY        = "npt_display"         # Negative pressure test chart display AI


class AdversarialWellheadIntegrityImageError(Exception):
    """Raised when Glyphward detects adversarial content in a subsea wellhead
    casing annulus pressure integrity AI rendered image above threshold 30.

    Consequence if not raised:
    - ANNULUS_PRESSURE: SCP build-up suppressed → barrier breach investigation
      not initiated → gas migration continues → potential blowout;
      Macondo 2010: 11 killed, 4.9M barrels released.
    - CONNECTOR_CAMERA: partial connector engagement classified as full →
      BOP-to-wellhead seal fails at high wellhead pressure → uncontrolled well.
    - SEAL_LEAK_CAMERA: methane bubble plume suppressed → seal assembly
      erosion continues → progressive barrier integrity loss.
    - NPT_DISPLAY: failed NPT classified as passed → mud displaced with
      seawater → no hydrostatic overbalance on compromised barrier →
      well control event; Macondo NPT misinterpretation parallel.
    Fail-safe: immediately read raw pressure transducer signals (independent of
    AI display); initiate well barrier investigation under NORSOK D-010 Section 4;
    do not proceed with mud displacement or completion operations until NPT is
    independently verified from raw sensor data; activate BOP if required.
    """

    def __init__(self, scan_id, score, context, well_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.well_id = well_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial wellhead integrity image: context={context.value} "
            f"score={score} well={well_id} scan_id={scan_id}"
        )


async def scan_wellhead_integrity_image(image_bytes, context, well_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"wellhead_integrity:{context.value}:{well_id}",
        "metadata": {
            "well_id": well_id,
            "context": context.value,
            "image_sha256": image_hash,
            "scan_timestamp_utc": datetime.now(timezone.utc).isoformat(),
        },
    }
    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()
    if result["score"] >= WELLHEAD_INTEGRITY_THRESHOLD:
        raise AdversarialWellheadIntegrityImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            well_id=well_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_wellhead_integrity_image before each wellhead AI classification call. On AdversarialWellheadIntegrityImageError for NPT_DISPLAY: immediately suspend well displacement operations; read raw drill string and kill line pressure transducer signals directly from the SCADA historian rather than the AI display; initiate a well barrier investigation under NORSOK D-010 Section 4.2; notify the Well Control Coordinator before proceeding with any activity that reduces hydrostatic overbalance above the primary well barrier. See also: offshore subsea pipeline riser AI prompt injection (related subsea safety AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions