Prompt injection in nuclear waste interim storage ISFSI dry cask AI

Four adversarial injection surfaces in nuclear waste ISFSI dry cask AI

1. Dry cask canister surface temperature trend display AI (Holtec SmartMonitor AI, NAC International IntelliStore AI, Orano DSS monitoring AI — ISFSI canister surface temperature trend monitoring AI)

Dry cask canister temperature monitoring — the measurement of canister surface temperatures using thermocouples welded to the outer surface of the stainless steel multipurpose canister (MPC) or the inner surface of the ventilated concrete overpack (VCO) — is the primary diagnostic indicator of dry cask heat removal performance. The NRC 10 CFR Part 72 licensing basis for each ISFSI dry storage system includes an approved maximum canister surface temperature (typically 300–400°C for the stainless steel MPC surface, depending on the storage system Certificate of Compliance) derived from the thermal analysis of the maximum heat load canister under the licensing basis ambient conditions. If the canister surface temperature exceeds the licensing basis maximum, the cladding temperature inside the canister — typically 50–150°C above the canister surface temperature, depending on the cask design — approaches or exceeds the NRC’s cladding temperature limits (400°C for normal conditions; higher for design basis events). At cladding temperatures above approximately 400°C in an inert atmosphere (the canister is backfilled with helium for heat conduction), zirconium alloy cladding creep rates increase significantly, reducing the cladding hoop strength and increasing the risk of cladding failure from internal fuel pellet swelling. The consequence of cladding failure in a sealed dry cask canister — the release of the fission product gas inventory from the failed fuel rods into the canister helium atmosphere — does not immediately produce a release to the environment (the sealed canister provides confinement), but it reduces the safety margin against subsequent canister breach under mechanical stress events. AI systems process rendered canister surface temperature trend display images — multi-channel strip chart renders of thermocouple readings from the monitoring channels on each dry cask — to classify thermal performance: normal (all channels within design envelope), marginal (one or more channels approaching licensing basis maximum), degraded (channel approaching or at licensing basis maximum — increased monitoring frequency and heat removal assessment required), and emergency (channel above licensing basis maximum — immediate NRC notification per 10 CFR 72.75 and emergency heat removal assessment required).

An adversarial perturbation on a rendered canister temperature trend display image that suppresses a rising thermocouple channel — applying a ±10 DN downward shift to the pixel region encoding the temperature trend line above the marginal threshold (reducing the apparent temperature from the marginal or degraded zone to within the normal design envelope) — causes the ISFSI monitoring AI to classify an actual canister overheating event as normal thermal performance, suppressing the increased monitoring frequency and heat removal assessment that a marginal classification requires. Canister overheating on an ISFSI can occur from: blocked air inlet or outlet vents (debris accumulation, ice formation, vegetation, or unauthorised obstruction); ambient temperature exceedances above the licensing basis maximum ambient (currently 38°C / 100°F for most US ISFSI licenses, which may be challenged by climate change — NUREG-2224 addresses climate change considerations for ISFSI thermal analyses); and heat load miscalculation at loading (canister loaded with higher decay heat assemblies than the approved loading curve permits). The consequence chain from undetected canister overheating to cladding damage is gradual — hours to days at the thermocouple reading rates used in ISFSI monitoring — making the temperature trend AI the primary early-warning tool in the monitoring regime, and adversarial suppression of a rising temperature trend the critical adversarial vulnerability.

2. Seismic event accelerometer array display AI (KINEMETRICS ETNA-2 AI, REFTEK 130 seismic AI, SYSCOM MS2003 strong-motion AI — ISFSI seismic monitoring system display AI)

ISFSI facilities are licensed under NRC 10 CFR Part 72 Appendix A with a site-specific seismic design basis: the Safe Shutdown Earthquake (SSE) ground motion level used in the dry cask structural design, typically specified as the 10,000-year return period ground motion at the ISFSI pad location. The dry cask storage systems (concrete overcpacks, storage modules, and MPC canisters) are qualified to withstand the SSE without loss of confinement (the MPC weld remains intact) or loss of shielding (the concrete overpack does not tip, slide, or fall). Seismic monitoring at ISFSI facilities — networks of strong-motion accelerometers installed on the ISFSI concrete pad, on selected dry casks, and at the ISFSI facility boundary — provide post-event acceleration records that are compared against the SSE design basis to determine whether the design basis has been exceeded and whether post-earthquake inspection is required per the ISFSI Technical Specifications. AI systems process rendered seismic accelerometer array display images — real-time accelerogram displays showing peak ground acceleration (PGA) at each accelerometer channel during and after a seismic event, or frequency-domain displays (response spectra) rendered by the seismic monitoring workstation — to classify seismic event consequence: below SSE (no inspection required beyond routine schedule), approaching SSE (enhanced inspection required within 24 hours), at or above SSE (immediate inspection of all casks and the ISFSI pad required per Technical Specifications, NRC notification required under 10 CFR 72.75 within 4 hours), and equipment malfunction (accelerometer channel out of range or offline — maintenance required before next inspectable event).

An adversarial perturbation on a rendered seismic accelerometer display image that suppresses a peak ground acceleration reading — applying a ±8 DN downward shift to the pixel region encoding the PGA bar or accelerogram trace at or approaching the SSE level (reducing the apparent PGA from the at-SSE or approaching-SSE range to the below-SSE range) — causes the seismic monitoring AI to classify an SSE-level seismic event as a routine below-SSE event, suppressing the immediate post-earthquake inspection and NRC notification that the Technical Specifications require when the SSE design basis is reached. The consequence of a suppressed SSE-level event classification depends on whether the seismic event caused structural damage to the dry casks or ISFSI pad: if cask tip-over or concrete overpack cracking occurred at the SSE design basis, the adversarially suppressed classification delays the inspection that would identify the damage before a secondary event (aftershock at reduced structural margin, or the cask in degraded condition remaining in storage). The Fukushima Daiichi spent fuel pool management during the 2011 Tōhoku earthquake — where the spent fuel pools remained cooled but required post-earthquake inspection to verify pool liner integrity — established the regulatory expectation for immediate post-earthquake inspection of radioactive fuel storage systems at the design basis seismic level. NRC RG 3.72 requires seismic monitoring at ISFSIs with appropriate sensitivity to detect design basis events — but does not specify adversarial robustness requirements for AI systems classifying rendered seismic monitoring display images.

3. Canister remote inspection camera AI (GE Inspection Technologies CARIS AI, Eddyfi Technologies dry cask inspection AI, Applus+ RTD remote canister camera AI — ISFSI dry cask visual inspection AI)

Remote visual inspection of dry cask canisters — using remotely operated camera systems deployed at the ISFSI pad to examine the accessible exterior surfaces of dry casks for physical damage, corrosion, seal degradation, or surface anomalies — is a component of the periodic inspection program required by 10 CFR Part 72 Technical Specifications and the ISFSI maintenance plan. For horizontal storage modules (NUHOMS DSC, Advanced NUHOMS), camera systems inspect the interior of the concrete horizontal storage module (HSM) to examine the stainless steel DSC (dry shielded canister) outer surface for signs of external corrosion, pitting, stress corrosion cracking (SCC), or mechanical damage. For vertical systems (Holtec HI-STORM, NAC MAGNASTOR), cameras inspect the annular air flow path between the MPC canister and the concrete overpack to detect canister surface anomalies and ventilation system obstructions. AI systems process rendered inspection camera images — digital video frames from radiation-hardened cameras operating in the high gamma dose rate environment of the ISFSI pad (typically 10–100 mSv/hr at contact with the concrete overpack surface) — to classify canister physical condition: normal (no surface anomalies beyond expected corrosion product discolouration), minor surface anomaly (corrosion product accumulation or minor surface pitting — enhanced monitoring required), significant surface anomaly (visible pitting, mechanical damage, or potential SCC initiation site — engineering evaluation and NRC notification required), and canister breach indication (visible perforation, crack, or seal degradation — emergency response per ISFSI Emergency Plan required).

An adversarial perturbation on a rendered canister inspection camera image that suppresses a surface anomaly indicator — applying a ±8 DN colour and texture shift to the pixel region encoding the corrosion pit, SCC crack, or mechanical damage feature on the canister surface (normalising the apparent surface condition to the uniform grey of intact stainless steel) — causes the inspection AI to classify a developing canister surface degradation as normal MPC exterior condition, suppressing the engineering evaluation and inspection interval reduction that a significant anomaly classification requires. The most significant emerging concern for dry cask canister integrity is chloride-induced stress corrosion cracking (Cl-SCC) of austenitic stainless steel MPC canisters stored in marine coastal environments (such as the Diablo Canyon ISFSI in California, the Surry ISFSI in Virginia, and multiple Japanese ISFSI sites in coastal locations): chloride deposition from marine aerosols on the outer MPC surface in the annular air flow gap, combined with the canister surface temperature and residual welding stresses, creates conditions under which SCC can initiate and propagate. EPRI (Electric Power Research Institute) research on Cl-SCC of SNF dry cask canisters (EPRI 3002005465, 2015) documented that Cl-SCC of austenitic stainless steel is possible at the canister surface temperatures and chloride concentrations present at coastal ISFSIs — making the canister inspection camera AI the primary surveillance tool for early SCC detection, and adversarial suppression of a surface anomaly indicator the critical adversarial gap.

4. ISFSI air inlet and outlet thermal imaging AI (FLIR Systems A615 thermal AI, Teledyne FLIR thermography AI, Jenoptik VarioCAM HD thermal AI — dry cask ventilation air flow thermal imaging AI)

Passive air cooling — the convective circulation of ambient air through the annular air path between the multipurpose canister (MPC) and the ventilated concrete overpack (VCO) or horizontal storage module (HSM) — is the primary and sole heat removal mechanism for dry cask storage systems. The passive air cooling design requires unobstructed air inlet vents at the base of the VCO or HSM and air outlet vents at the top, with a minimum annular flow path cross-section throughout the vertical height of the overpack. If the air inlet or outlet vents become obstructed — by debris accumulation (leaf litter, bird nests, ice formation from freezing precipitation), unauthorised physical obstruction (the ISFSI security concern addressed in NRC Regulatory Issue Summary 2012-05), or structural damage following a seismic or extreme weather event — the passive air cooling circuit is degraded or interrupted, and canister surface temperatures begin to rise at a rate determined by the degree of obstruction and the canister decay heat load. Thermal imaging cameras mounted at the ISFSI to monitor air inlet and outlet vent temperatures — imaging the air column exiting the outlet vents and comparing to the inlet air temperature — provide an integrated measure of heat removal performance: the outlet air temperature above the ambient inlet temperature is proportional to the heat removal rate, and a decrease in outlet-to-inlet temperature differential indicates reduced air flow (ventilation degradation). AI systems process rendered thermal camera images of the dry cask air inlet and outlet zones — false-colour thermographic images of the VCO or HSM vent arrays — to classify ventilation performance: normal (outlet temperature above inlet by design-basis value × 0.9–1.1 margin), marginal (reduced temperature differential — vent inspection required), degraded (significantly reduced differential — immediate vent obstruction investigation and clearance required), and blocked (no temperature differential — emergency heat removal assessment, potential for forced ventilation or water addition per ISFSI Emergency Plan).

An adversarial perturbation on a rendered thermal camera image of the dry cask air outlet vents that suppresses the temperature differential reduction — applying a ±10 DN upward shift to the pixel values encoding the outlet vent air column thermal image (increasing the apparent outlet temperature above the reduced actual outlet temperature to the expected design-basis differential range) — causes the ISFSI ventilation monitoring AI to classify a degraded or blocked passive cooling system as operating normally, suppressing the vent inspection and obstruction clearance that the degraded performance requires. As the blocked ventilation condition persists undetected, canister surface temperatures rise on the timescale of hours to days (depending on cask thermal inertia and decay heat load), eventually reaching the licensing basis temperature limits. For casks in high-burnup service or with newer shorter-cooled fuel, the decay heat load is higher and the time from blocked ventilation to licensing basis temperature exceedance is shorter — in some high-burnup cask designs, complete ventilation blockage with high-heat-load canisters could result in licensing basis temperature exceedance within 24–72 hours. NRC NUREG-2224 (Dry Storage and Transportation of High-Burnup Spent Nuclear Fuel) documents the thermal sensitivity of high-burnup fuel to ventilation degradation as an emerging concern that requires monitoring vigilance — precisely the vigilance that adversarial suppression of ventilation thermal AI removes.

Integration: ISFSI dry cask AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for ISFSI dry cask AI belongs at every rendered-image ingestion boundary in the dry cask safety monitoring pipeline — before canister surface temperature trend AI processes rendered thermocouple display images, before seismic event accelerometer display AI processes rendered accelerogram or PGA display images, before canister inspection camera AI processes rendered inspection camera frames, and before air inlet/outlet thermal imaging AI processes rendered thermographic images. Threshold 30 reflects the radiological consequence of cladding failure (internal fission product release to canister atmosphere) combined with the multi-decade ISFSI storage commitment (US dry cask storage timelines now extend to 100+ years per NUREG-2157) and the remote, infrequently monitored nature of ISFSI operations.

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"

# Nuclear waste ISFSI dry cask AI contexts: threshold 30
# NRC 10 CFR Part 72 (ISFSI licensing basis);
# 10 CFR Part 72 Appendix A (General Criteria for ISFSIs);
# NUREG-1567 (Standard Review Plan for Dry Storage Facilities);
# NUREG-2224 (High-Burnup SNF Dry Storage and Transportation).
ISFSI_DRY_CASK_THRESHOLD = 30


class ISFSIDryCaskAIContext(Enum):
    CANISTER_TEMPERATURE = "canister_temperature"  # Surface temperature trend AI
    SEISMIC_ACCELEROMETER = "seismic_accelerometer" # PGA display / accelerogram AI
    CANISTER_INSPECTION   = "canister_inspection"   # Remote visual inspection AI
    AIR_THERMAL_IMAGING   = "air_thermal_imaging"   # Inlet/outlet ventilation AI


class AdversarialISFSIDryCaskImageError(Exception):
    """Raised when Glyphward detects adversarial content in an ISFSI dry cask
    AI rendered monitoring image above threshold 30.

    Consequence if not raised:
    - CANISTER_TEMPERATURE: overheating suppressed → cladding temperature
      above NRC licensing basis limit → Zr alloy creep failure → fission
      product gas release into canister helium atmosphere → cladding damage
      without external indication; radiological consequence if subsequent
      canister breach.
    - SEISMIC_ACCELEROMETER: SSE-level PGA suppressed → post-earthquake
      inspection not initiated → structural damage to cask or pad
      undetected → degraded margin for aftershock or continued storage.
    - CANISTER_INSPECTION: Cl-SCC crack suppressed → stress corrosion
      cracking propagation undetected → canister breach → radiological
      release to ISFSI environment; coastal ISFSI Cl-SCC risk.
    - AIR_THERMAL_IMAGING: ventilation blockage suppressed → passive
      cooling degraded → canister temperature rise to licensing basis
      limit over 24-72 hrs (high-burnup fuel); NUREG-2224 scenario.
    Fail-safe: halt AI monitoring classification; conduct manual
    thermocouple readout and physical vent inspection per 10 CFR Part
    72 Technical Specifications before resuming AI monitoring.
    """

    def __init__(self, scan_id: str, score: int,
                 context: ISFSIDryCaskAIContext,
                 isfsi_id: str, cask_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.isfsi_id = isfsi_id
        self.cask_id = cask_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial ISFSI dry cask image: "
            f"context={context.value} score={score} "
            f"isfsi={isfsi_id} cask={cask_id} scan_id={scan_id}"
        )


async def scan_isfsi_dry_cask_image(
    image_bytes: bytes,
    context: ISFSIDryCaskAIContext,
    isfsi_id: str,
    cask_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an ISFSI dry cask AI rendered monitoring image for adversarial content.

    Fail-safe contract: AdversarialISFSIDryCaskImageError or httpx error →
    halt ISFSI AI monitoring classification for affected cask/zone; conduct
    manual thermocouple readout (Technical Specification surveillance) and
    physical vent/surface inspection before resuming AI-driven temperature
    trending or ventilation monitoring; notify Radiation Protection Manager
    and NRC resident inspector per 10 CFR 72.75 if cask condition uncertain.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"isfsi_dry_cask:{context.value}:{isfsi_id}:{cask_id}",
        "metadata": {
            "isfsi_id": isfsi_id,
            "cask_id": cask_id,
            "context": context.value,
            "image_sha256": image_hash,
        },
    }
    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"] > ISFSI_DRY_CASK_THRESHOLD:
        raise AdversarialISFSIDryCaskImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            isfsi_id=isfsi_id,
            cask_id=cask_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_isfsi_dry_cask_image at each ISFSI monitoring AI rendered-image ingestion boundary: before canister surface temperature trend AI (threshold 30), before seismic event accelerometer display AI (threshold 30), before canister remote inspection camera AI (threshold 30), and before air inlet/outlet thermal imaging AI (threshold 30). On AdversarialISFSIDryCaskImageError for CANISTER_TEMPERATURE context: immediately conduct independent thermocouple readout using the backup manual readout channel (required by 10 CFR Part 72 Technical Specification surveillance procedure), notify Radiation Protection Manager, and assess whether 10 CFR 72.75 four-hour NRC notification is required before resuming AI temperature trending. See also: nuclear power plant digital I&C AI prompt injection (related nuclear safety I&C adversarial context) and geotechnical slope monitoring AI prompt injection (related long-term infrastructure monitoring AI adversarial gap). Get early access

Related questions