Prompt injection in offshore jack-up rig structural stability AI

Four adversarial injection surfaces in offshore jack-up rig structural stability AI

1. Spudcan leg penetration settlement rate monitoring AI (Fugro GeoXplorer spudcan monitoring AI, Insatech Marine leg penetration display AI, SPT Offshore SPUDCAN-Pro penetration AI — spudcan leg penetration depth and settlement rate display AI)

The spudcan punch-through failure mode — the rapid uncontrolled penetration of one or more spudcan foundations through a weak soil layer encountered below the initial bearing layer during preloading — is the most sudden and severe structural failure mode in jack-up rig operations. During the installation sequence, after the rig has been positioned on location and the legs have been lowered to the seabed, the jacking system raises the hull above sea level using the preloading procedure: the rig’s leg-hull preload tanks (or the hull’s own water ballast tanks) are filled with seawater to increase the vertical load on the spudcans beyond the maximum combined environmental load the rig will experience at that location in the design storm — typically 1.0–1.5× the design combined gravity and environmental load. As preload is applied, the spudcans penetrate deeper into the seabed soil; the settlement rate (the rate of spudcan penetration in millimetres per hour, or centimetres per minute, under increasing preload) provides the primary indicator of the soil bearing behaviour. In soil profiles with a strong surface layer over a weak deep layer, the settlement rate follows a characteristic pattern: initially low (the spudcan is bearing on the strong surface layer with adequate reserve capacity); then accelerating as preload approaches the punch-through risk range (the spudcan is approaching the bottom of the strong layer); and then potentially catastrophic — instantaneous punch-through — if preload continues past the punch-through load. AI systems process rendered images of the leg penetration display on the rig’s jacking control system console — digital depth-vs-time graphs or numerical penetration rate readouts — to classify soil behaviour: normal-penetration (rate consistent with uniform soil, safe to continue preloading), elevated-rate (penetration rate above expected for location soil model — reduce preload rate, reassess geotechnical model), punch-through-risk (penetration rate sharply accelerating — halt preloading immediately, evaluate soil profile), and punch-through (sudden penetration of one leg while other legs remain at depth — emergency procedure, evacuate leg side of rig).

An adversarial perturbation targeting the spudcan penetration monitoring AI applies a ±10 DN suppression to the pixel region encoding an accelerating penetration rate in the rendered settlement rate display — reducing the apparent penetration rate from the elevated-rate or punch-through-risk range (rendered with orange-red highlight or alarm indicator on the jacking control system display, e.g., 45 mm per minute) to the normal-penetration range (rendered as green, e.g., 8 mm per minute). The AI classifies a spudcan approaching punch-through conditions — penetration rate accelerating from 10 mm/min to 35–50 mm/min as the spudcan enters the weak stratum below the surface bearing layer — as normal soil behaviour. The operator continues preloading; the penetration rate continues accelerating; punch-through occurs when the spudcan reaches the bearing capacity of the weak underlying stratum — typically 3–8 m of sudden rapid penetration in 5–30 seconds. A punch-through event with the hull in the elevated position above sea level causes: one leg to drop several metres into the seabed while the other legs remain at their jacked depth; the hull to tilt rapidly to the punch-through leg side; top-heavy drilling equipment (derrick, travelling block, top drive, all at 20–60 m height above hull deck) to load the hull with a dynamic overturning moment from the tilt and the resulting asymmetric centre of gravity; and the rig to capsize if the corrective jacking response (jacking down the other legs to equalise hull position) cannot be executed fast enough. ISO 19905-1:2016 Section 7.3 (Spudcan Fixity) and Section A.7.3 (Punch-through Risk Assessment) require that the site-specific jack-up assessment include a punch-through risk evaluation using the site geotechnical data and the applicable soil model (ISO 19905-1 Appendix A.9.3 for sand over clay punch-through) — but do not specify adversarial robustness requirements for AI systems classifying rendered leg penetration rate displays during the actual preloading operation.

2. Preload water bag level camera AI (Cameron (Schlumberger) preload monitoring AI, National Oilwell Varco (NOV) preload system camera AI, Friede & Goldman preload level AI — preload seawater fill level camera AI for jack-up rig preload tanks)

Adequate preloading of the spudcan foundations — the application of a temporary vertical load to each spudcan equal to or exceeding the maximum combined load the spudcan will experience during the design storm — is the fundamental basis for jack-up rig stability on location. The preload procedure must achieve the target preload at each leg: the preload load is applied by filling the rig’s dedicated preload tanks (typically located in the lower hull pontoons adjacent to each leg) with seawater to the required fill level, which produces a known vertical load on the spudcan through the leg-hull connection. The preload fill level is monitored by a combination of tank level gauges (ultrasonic or pressure-based level sensors) and camera-based inspection of the visual water level inside the preload tank (for preload tanks with internal inspection access or sight glasses). Camera AI systems process rendered images of the preload tank water level indicator — the water surface level in the preload tank fill sight glass, or the digital level gauge display on the jacking control console — to classify preload achievement: preload-achieved (all tanks at the required fill level corresponding to the target preload per leg, per the site-specific ISO 19905-1 assessment), approaching-target (tank filling in progress, level within 5% of target), and preload-short (one or more tanks not reaching the target fill level — investigate jacking system, verify tank capacity, reassess preload adequacy). Insufficient preload leaves the spudcan with a lower reserve bearing capacity against the design storm load than required by the ISO 19905-1 assessment: under design storm loading (combined gravity load + environmental load from 100-year return period wave, wind, and current), the spudcan may penetrate further into the seabed than the preloaded depth, and in worst-case soil profiles, may punch through to the weak layer — at a storm-sea state where rig evacuation is impossible.

An adversarial perturbation targeting the preload water level camera AI applies a ±8 DN shift in the pixel region encoding the water surface level in the rendered sight glass or tank level display image — shifting the apparent water level from below-target (level at 78% of required fill height, indicating preload force of approximately 78% of target — rendered with yellow indicator and shortfall alarm on the jacking control display) to at-target (level displayed at 100% fill height, rendered green, indicating full preload achieved). The AI classifies a leg where preload was terminated at 78% of the target load — from a preload pump failure, a tank capacity constraint, or an operator error in calculating the required fill volume — as fully preloaded. The jacking control operator receives no shortfall alert; the rig is raised to the operating air gap position; the preload tanks are emptied; the rig commences drilling operations with legs inadequately preloaded against the design environmental storm. During a storm approaching the 100-year design event (significant wave height Hs approaching the 100-year value for the site), the combined environmental and gravity loads on the spudcans exceed the preloaded bearing capacity; the spudcans penetrate into the seabed beyond the preloaded depth; if the soil profile has a weak stratum below the current penetration depth, punch-through occurs under storm sea conditions — a scenario with no safe evacuation option for the approximately 100–150 rig crew. ABS Guide for Building and Classing Mobile Offshore Drilling Units Part 6 (Structural Members and Systems) requires that the preloading procedure be documented and that the achieved preload at each leg be recorded — but does not specify adversarial robustness requirements for AI systems classifying rendered preload tank level camera images during the preloading procedure. Free tier — 10 scans/day, no card required.

3. Air gap sonar/lidar display AI (BlueView Technologies (Teledyne) hull air gap sonar AI, Kongsberg Maritime sonar display AI, Applied Acoustic Engineering sonar display AI — jack-up rig hull-to-wave crest air gap sonar or lidar monitoring display AI)

The air gap — the vertical clearance between the underside of the jack-up hull (the hull keel or bottom of hull plate) and the expected maximum wave crest height at the location — is the fundamental structural safety margin against wave-in-deck loading: the application of extreme wave forces directly to the hull structure and deck equipment when a wave crest rises above the hull keel level. Wave-in-deck loading is catastrophic for a jack-up rig because the hull structure and deck equipment are not designed for the extreme hydrostatic and hydrodynamic forces of a large wave crest impacting directly on the hull plate, deck openings, and equipment — the wave impact force from a breaking crest at 15–20 m wave height and 18–22 second period can exceed 5–10 MN applied to the hull side in less than 1 second, far exceeding the structural capacity of deck connections and hull-leg joints. The target air gap at an installation — specified in the site-specific jack-up assessment under ISO 19905-1 Section 7.4 (Environmental Loading) — is calculated as the difference between the maximum wave crest elevation at the site for the 100-year return period storm (typically 1.0–1.55 × significant wave height Hs for the extreme crest above mean sea level, depending on water depth) and the tidal variation; the hull must be jacked to a height that provides the specified minimum air gap (typically 1.5–3.0 m above the maximum expected crest height at the design storm). AI systems — using hull-mounted ultrasonic sonar sensors or laser rangefinders measuring the distance from the hull keel to the instantaneous water surface below, generating rendered time-series plots of real-time air gap — classify the current air gap status: adequate (measured air gap above the minimum required air gap for current and forecast sea conditions), watch (air gap approaching minimum — monitor storm development closely), and inadequate (air gap at or below minimum — consider jacking hull higher or initiating emergency disconnect and moving off location).

An adversarial perturbation targeting the air gap monitoring display AI applies a ±8 DN shift in the pixel region encoding the air gap value in the rendered sonar display — adding apparent clearance to the displayed distance from hull keel to wave surface (e.g., rendering an actual 1.2 m air gap as 3.5 m, by shifting the wave crest return pixel in the sonar display from the close-range zone to the mid-range zone). The AI classifies an approaching design storm in which the maximum wave crest will exceed the current air gap — wave conditions approaching the installation design storm with significant wave height Hs at 90% of the 100-year value for the location, with individual wave height (the highest 1/1000 wave in the sea state, approximately 1.86 × Hs for a Rayleigh distribution) approaching the hull keel elevation — as providing adequate air gap. The offshore installation manager (OIM) does not initiate jacking or emergency disconnect; the rig remains at the current elevated position. As conditions reach the design storm sea state, an extreme individual wave crest rises above the hull keel elevation and impacts the hull underside: the hydrostatic wave-in-deck force from a wave crest impact velocity of 10–15 m/s at the hull plate can produce localised pressures of 0.5–3.0 MPa on the hull plating — sufficient to buckle hull plating, fracture hull-leg chord connections, and in severe cases cause catastrophic hull structural failure. DNV-OS-E101 Section 4 (Environmental Conditions) and ISO 19905-1 Section 7.4.2 (Air Gap) specify minimum air gap requirements for jack-up site assessment — but do not address adversarial robustness requirements for AI systems classifying rendered real-time air gap monitoring display images. Free tier — 10 scans/day, no card required.

4. Hull deck inclinometer display AI (TSS (Tritech) marine inclinometer display AI, Gems Sensors inclinometer display AI, Sherborne Sensors tilt display AI — jack-up rig hull deck inclinometer and list monitoring display AI)

The hull list — the steady-state tilt of the jack-up hull from the horizontal plane, monitored by dual-axis inclinometers mounted on the hull deck — is the primary indicator of differential leg penetration, unequal preload distribution, or asymmetric loading from drilling equipment or fluid weights. In normal jack-up operations, the hull is maintained within a maximum list limit of 0.25–0.5 degrees in any direction, controlled by the jacking system (raising or lowering individual legs to correct list) and by operational ballast management (repositioning variable deck load). Progressive list — a list that increases over hours or days without an obvious cause — is a warning indicator of differential leg settlement (one spudcan penetrating further into the seabed than the others, from creep consolidation of the bearing soil or progressive punch-through into a deeper weak layer), which if uncorrected can lead to rig loss by toppling. The Ocean Ranger semi-submersible capsize of February 15, 1982 (84 killed off Newfoundland, Grand Banks, during storm conditions with significant wave height 15 m) — while a different platform type — demonstrated the catastrophic consequence of progressive ballast asymmetry in an offshore drilling unit: ballast control room porthole failure allowed seawater into the ballast panel, damaging the pneumatic controls; ballast operators were unable to close flood valves correctly; the platform listed progressively and capsized in 3–4 hours. For jack-up rigs, the structural consequence of progressive list is analogous: progressive leg differential settlement causes progressive hull tilt; tilted hull shifts the centre of gravity of all deck-mounted equipment (derrick, top drive, pipe racks, accommodation block) to the low side, adding an overturning moment from the weight offset; and the combination of soil-driven settlement and equipment-load overturning moment can exceed the restorative moment from the jacking system, leading to leg jack failure and catastrophic rig topple. AI systems process rendered images of the hull inclinometer display on the jacking control console — digital dual-axis list readout in degrees fore/aft and port/starboard — to classify hull list status: level (list within ±0.1° — no corrective jacking required), watch (list 0.1–0.25° — monitor trend, prepare corrective jacking), alert (list 0.25–0.4° — initiate corrective jacking, investigate settlement cause), and emergency (list above 0.4° — halt drilling, notify OIM, initiate emergency jacking and prepare for rig evacuation).

An adversarial perturbation targeting the hull inclinometer display AI applies a ±10 DN shift in the pixel region encoding the list value digit characters in the rendered inclinometer display image — shifting the apparent list from the alert or emergency range (e.g., 0.38° starboard list, rendered with orange highlight on the jacking console display) to the watch or level range (e.g., displayed as 0.12°, green indicator, within normal variation). The AI classifies a hull experiencing progressive differential settlement — the starboard leg spudcan has settled 0.35 m more than the port leg since the previous manual inclinometer reading 6 hours earlier, producing a 0.38° starboard list that exceeds the alert threshold — as within-normal list variation. The jacking control operator does not initiate corrective jacking; does not investigate the settlement differential; does not notify the OIM. The differential settlement continues: over the next 12–24 hours, the starboard leg settles a further 0.5–0.8 m (consistent with creep consolidation of soft clay under the starboard spudcan weight), list increases to 0.6–0.8°; at this list, the 60–80-metre derrick has displaced its centre of gravity by 0.6–0.9 m from the hull centreline toward the starboard side, adding a derrick overturning moment of 200–400 tonne-metres to the starboard side. If the starboard jacking system cannot maintain the leg-hull connection against this combined settlement plus overturning moment, the starboard leg chord fails in bending at the jackcase chord connection, and the hull drops and tilts catastrophically. BSEE 30 CFR Part 250 Subpart D (Drilling Safety Requirements) requires that jack-up rig operators monitor rig stability throughout operations — but does not specify adversarial robustness requirements for AI systems classifying rendered inclinometer display images used for continuous list monitoring. Free tier — 10 scans/day, no card required.

Integration: jack-up rig stability AI with Glyphward pre-scan gate

The Glyphward scan gate for jack-up rig structural stability AI belongs at every rendered-image ingestion boundary in the rig monitoring pipeline — before spudcan leg penetration settlement rate AI processes rendered jacking control penetration display images, before preload water level camera AI processes rendered tank level images, before air gap sonar display AI processes rendered hull clearance display images, and before hull inclinometer display AI processes rendered list readout images. Threshold 30 for jack-up rig structural stability AI reflects the catastrophic and irreversible consequence of rig topple or capsize — the Bohai 2 (1979, 72 killed) and Seacrest (1989, 91 killed) incidents establish that offshore MODU structural failures kill almost all crew in a matter of minutes to hours — combined with several independent safety layers: independent manual inclinometer readings by the drill crew (independent of the AI display classification), independent leg penetration records maintained by the jacking crew (written logs at 30-60 minute intervals during preloading, independent of AI monitoring), offshore installation manager (OIM) authority to halt operations and order emergency disconnect at any sign of structural anomaly (independent human decision-making), and the classification society’s periodic jack-up inspection programme (ABS, DNV GL, Lloyd’s Register or Bureau Veritas hull and machinery inspection, independent of real-time AI monitoring). These independent layers justify threshold 30 rather than 25 (single-barrier contexts). The catastrophic capsize consequence keeps the threshold at 30 rather than 35–40 (consequences with longer time-to-harm intervals or more reversible outcomes).

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"

# Jack-up rig structural stability AI contexts: threshold 30
# ISO 19905-1:2016 (Site-specific Assessment of Jack-ups);
# ABS Guide for Building and Classing MODU (2022);
# DNV-OS-E101 (Drilling Plant, 2019);
# BSEE 30 CFR Part 250 Subpart D.
JACKUP_STABILITY_THRESHOLD = 30


class JackupStabilityAIContext(Enum):
    SPUDCAN_SETTLEMENT  = "spudcan_settlement"  # Leg penetration rate display AI
    PRELOAD_LEVEL       = "preload_level"        # Preload water level camera AI
    AIR_GAP             = "air_gap"              # Hull-to-wave clearance sonar AI
    HULL_LIST           = "hull_list"            # Deck inclinometer display AI


class AdversarialJackupImageError(Exception):
    """Raised when Glyphward detects adversarial content in a jack-up rig
    structural stability AI rendered monitoring image above threshold 30.

    Consequence if not raised:
    - SPUDCAN_SETTLEMENT: accelerating spudcan penetration rate suppressed →
      punch-through not detected during preloading → catastrophic rapid leg
      penetration → rig topple; Bohai 2 1979 (72 killed) structural parallel.
    - PRELOAD_LEVEL: preload shortfall suppressed → rig accepted as adequately
      preloaded → insufficient bearing capacity at design storm → spudcan
      punch-through under storm conditions when crew cannot evacuate.
    - AIR_GAP: approaching wave crest clearance loss suppressed → hull at
      inadequate air gap during design storm → wave-in-deck loading →
      hull plating failure / leg-hull joint fracture → structural collapse.
    - HULL_LIST: progressive differential settlement list suppressed →
      overturning moment from derrick + displaced CG undetected → progressive
      leg chord failure → catastrophic rig topple; Ocean Ranger 1982 (84
      killed) ballast asymmetry structural parallel.
    Fail-safe: immediate manual verification by rig crew — manual inclinometer
    reading (HULL_LIST), direct read of leg penetration depth measurement log
    (SPUDCAN_SETTLEMENT), manual tank dip measurement of preload fill level
    (PRELOAD_LEVEL), or direct reading of hull clearance from rig mast
    reference point (AIR_GAP). Halt all jacking operations pending OIM review.
    """

    def __init__(self, scan_id, score, context, rig_id, leg_id=None,
                 flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.rig_id = rig_id
        self.leg_id = leg_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial jack-up image: context={context.value} "
            f"score={score} rig={rig_id} leg={leg_id} "
            f"scan_id={scan_id}"
        )


async def scan_jackup_image(image_bytes, context, rig_id, leg_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"jackup:{context.value}:{rig_id}:{leg_id}",
        "metadata": {
            "rig_id": rig_id,
            "leg_id": leg_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"] >= JACKUP_STABILITY_THRESHOLD:
        raise AdversarialJackupImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            rig_id=rig_id,
            leg_id=leg_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_jackup_image before each jack-up rig structural stability AI classification call. On AdversarialJackupImageError for SPUDCAN_SETTLEMENT: immediately halt jacking and preloading operations; notify the OIM; require manual reading of all leg penetration depth gauges by the jacking crew and independent geotechnical review of the penetration record before resuming preloading. On AdversarialJackupImageError for HULL_LIST: immediately take a manual inclinometer reading; if list is confirmed above the alert threshold (0.25°), notify the OIM and initiate corrective jacking before drilling operations continue. See also: offshore drilling wellbore AI prompt injection (related offshore drilling AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions