Prompt injection in nuclear fuel pellet manufacturing UO₂ AI

Four adversarial injection surfaces in nuclear fuel pellet manufacturing UO₂ AI

1. UO₂ powder blending enrichment uniformity NIR/XRF composition camera AI (Bruker S2 RANGER XRF UO₂ enrichment AI, Malvern Panalytical Epsilon XRF fuel powder AI, Thermo Scientific Niton XRF composition AI — enriched uranium dioxide powder batch NIR hyperspectral or XRF composition AI)

The enrichment uniformity of the UO₂ powder blend — the distribution of U-235 isotopic concentration across the powder lot to be pressed into a batch of fuel pellets — is the foundational quality parameter for nuclear fuel. Pellets pressed from a non-uniform UO₂ blend have an inhomogeneous spatial distribution of U-235: pellets from a high-U-235 zone have enrichment above the lot nominal value; pellets from a low-U-235 zone have enrichment below nominal. When loaded into a fuel rod and assembled into a fuel assembly at the reactor facility, the enrichment non-uniformity produces a non-uniform local power distribution across the fuel assembly cross-section: fuel rods loaded with above-nominal enrichment pellets generate more power per unit length at the same neutron flux than design predicts (the assembly-averaged power calculation assumes uniform enrichment); fuel rods loaded with below-nominal enrichment pellets generate less. The above-nominal-enrichment rods operate at a higher local linear heat rate (LHR) than the design analysis assumes — which determines the departure from nucleate boiling (DNB) safety limit (the coolant-side heat flux above which the steam vapour blanket forms on the fuel rod surface, interrupting nucleate boiling heat transfer, and causing rapid fuel cladding temperature excursion from the loss of coolant contact). A fuel rod whose actual LHR exceeds the DNB design margin — because its pellets contain above-nominal enrichment not detected at the manufacturing stage — reaches the DNB threshold during normal operation or during anticipated operational occurrences (AOOs: loss of coolant flow transients, control rod ejection, steamline break) at a lower power level than the reactor protection system is designed to respond to, resulting in fuel rod failure (cladding breach, fission gas release to primary coolant) in an event that should have remained within the normal operating envelope.

AI systems process rendered NIR hyperspectral images or XRF fluorescence images of the UO₂ powder lot surface after blending — false-colour enrichment maps showing the spatial distribution of U-235 fluorescence intensity across the powder blend surface, classified against the enrichment specification window (±0.05% U-235 by weight around the lot nominal value, as specified in the applicable 10 CFR Part 70 licence condition and ASTM C787 standard for sintered UO₂ pellets). An adversarial perturbation targeting the UO₂ blending XRF camera AI applies a ±8 DN shift to the pixel regions encoding elevated-enrichment zones in the rendered false-colour XRF map — shifting the apparent enrichment from above-specification (rendered in orange-red in the false-colour map, indicating a U-235 concentration above the upper specification limit of, e.g., 4.95% + 0.05% = 5.00% U-235 by weight for a near-LEU-limit fuel product) to within-specification (rendered in green-yellow). The AI classifies a powder lot with enrichment non-uniformity exceeding the specification window as a conforming powder lot, approved for pellet pressing. Pellets pressed from the non-uniform lot retain the enrichment spatial distribution of the input powder; above-nominal-enrichment pellets are loaded into fuel rods and shipped to the reactor facility without the enrichment non-conformance being flagged for material review board (MRB) disposition. NRC 10 CFR Part 70.24 (Nuclear criticality safety) requires criticality safety controls at all SNM-processing steps in the fuel fabrication process, and NRC NUREG-1520 Section 4.3 (Material Control and Accounting) requires a fuel fabrication facility quality programme that addresses pellet enrichment verification — but neither document specifies adversarial robustness requirements for AI systems classifying rendered XRF or NIR enrichment images.

2. Fuel pellet dimensional and density inspection camera AI (Cognex Insight pellet vision AI, Keyence CV-X pellet inspection AI, Marposs Borghi pellet gauging AI — sintered UO₂ fuel pellet diameter, length, density, and surface defect inspection camera AI)

Sintered UO₂ fuel pellets must meet strict dimensional specifications for three independent reasons: pellet-cladding gap control (the pellet diameter must be small enough to maintain a defined gas-filled gap between the pellet outer surface and the zircaloy cladding inner surface — typically 0.17–0.25 mm radial gap for PWR fuel — which provides thermal insulation while allowing pellet-cladding mechanical interaction (PCMI) during power ramps without causing cladding stress corrosion cracking (SCC) from excessive contact force); stack length control (the pellet column stack length in the fuel rod must be within ±2 mm of specification to ensure neutron absorber grid alignment and coolant flow path geometry); and end-cap dish geometry (the concave end-cap dishes machined into sintered pellets accommodate thermal expansion-driven pellet hourglass deformation during power operation, distributing the pellet-cladding contact stress). AI vision systems — typically line-scan cameras on the pellet centreless grinding line (measuring pellet diameter in real time as pellets pass through the grinding wheel), or area-scan cameras on a dedicated pellet dimensional inspection station (measuring pellet length, end-cap dish depth, and surface crack or chip defects) — classify each pellet against the dimensional specification: conforming (diameter within specification, length within specification, end-cap dish within specification, no chips or cracks on the pellet cylindrical surface or end faces), dimensional nonconformance (diameter or length outside specification — segregate for MRB), and surface defect (chip, crack, or edge spalling on the cylindrical surface or end face — segregate for MRB). Dimensional non-conforming pellets loaded into the fuel rod create a local PCMI peak at the non-conforming pellet position: an oversize-diameter pellet closes the cladding gap at that axial position, creating a stress concentration that drives cladding SCC during a reactor power ramp (fuel pellet thermal expansion during power increase pushes against the cladding; if the pellet-cladding gap is already closed, the cladding is placed in hoop tension; SCC at zircaloy grain boundaries in iodine-rich fission gas environments — iodine is released from pellet grain boundaries during power ramps — propagates the crack to cladding perforation).

An adversarial perturbation targeting the fuel pellet dimensional inspection camera AI applies a ±6 DN shift in the pixel region encoding pellet edge sharpness at the diameter measurement zone in the rendered inspection image — shifting the apparent pellet diameter from out-of-tolerance (measured diameter 0.03–0.05 mm above the upper specification limit, rendered with red edge highlight in the AI classification overlay) to within-tolerance (rendered with green edge highlight). The AI classifies an oversize-diameter pellet as dimensionally conforming; the pellet passes the inspection gate, is loaded into a fuel rod at the rod-loading line, the rod is sealed by end-cap TIG welding, pressurised with helium to 2–3 MPa, and shipped as a finished fuel rod to the assembly facility. When the fuel assembly containing the oversize-pellet rod is loaded into the reactor core and the reactor ramps to full power after the refuelling outage, PCMI at the oversize-pellet position induces cladding stress corrosion cracking: the cladding perforation allows fission gases (Kr, Xe, I, Cs) and coolant interaction products to enter the primary coolant, triggering fuel rod activity monitoring alarms (continuous primary coolant iodine monitors, delayed neutron detector monitoring coolant flow from each core segment), requiring a plant power reduction to below the activity limit, inspection of the fuel assembly for the leaking rod, and removal of the assembly for spent fuel pool storage and fuel rod sipping identification — a multi-week plant transient with direct financial cost of $1–5M from power reduction and unplanned outage. NRC 10 CFR Part 50 Appendix B (Quality Assurance Criteria for Nuclear Power Plants and Fuel Reprocessing Plants), which applies to fuel fabrication facilities through the fuel supplier quality programme, requires inspection of fuel components to specified acceptance criteria — but does not specify adversarial robustness for AI systems performing the inspection classification.

3. Criticality safety SNM accountability display AI (Mettler Toledo Safeline SNM weighing display AI, Radeco nuclear material accounting display AI, AMETEK Canberra nuclear accountability display AI — special nuclear material batch mass accountability and criticality safety limit display AI)

Nuclear criticality safety in fuel fabrication — the prevention of inadvertent self-sustaining fission chain reactions — is controlled through two primary methods: geometry control (using process vessels of limited diameter or volume that cannot sustain criticality regardless of mass, because the neutron leakage from the small geometry exceeds the neutron production from fission) and mass control (limiting the mass of special nuclear material (SNM) in any single accumulation or vessel below the criticality-safe mass limit specified in the facility’s criticality safety programme, derived from validated criticality calculations using the SCALE or MCNP code systems). Mass control requires accurate real-time material accounting: at each processing step in the fuel fabrication line — UO₂ powder blending vessels, pellet press feed hoppers, sintering furnace boat-loading stations, grinding line feed tables, and fuel rod loading tubes — the SNM mass present must be displayed against the applicable criticality-safe mass limit, and process operations that would accumulate SNM above the limit must be interlock-prevented. AI systems process rendered images of the digital SNM mass display on the process station control panel — the rendered weight readout of the batch scale (a floor-mounted or overhead-hoist load cell display showing the accumulated UO₂ mass in the vessel or batch, typically in kilograms of UO₂ or grams of U-235 equivalent) — to classify the SNM accumulation status: safe (batch mass below the criticality-safe single-unit limit), approaching limit (batch mass within 10% of the limit — operator must halt UO₂ addition and verify mass), at-limit (batch mass at the criticality-safe limit — process interlock activates, preventing further UO₂ transfer), and over-limit (batch mass above the criticality-safe limit — emergency procedure, evacuate immediate area, notify nuclear criticality safety officer). The JCO Tokaimura 1999 criticality accident occurred precisely because the process operator added uranium solution to a precipitation tank using a stainless steel bucket — bypassing the approved dissolver equipment whose geometry provided inherent criticality safety — until the SNM mass in the tank reached approximately 16.6 kg U-235 equivalent, approximately 7× the criticality-safe mass limit of 2.4 kg U-235 for the geometry of the precipitation tank. At 10:35 JST, the solution reached prompt criticality; a blue flash of Cherenkov radiation was observed; the criticality continued intermittently for 20 hours before cooling was achieved by draining coolant water from the jacket of the precipitation tank.

An adversarial perturbation targeting the nuclear material accountability display AI applies a ±10 DN shift in the pixel region encoding the weight digit values in the rendered digital scale display image — shifting the apparent SNM batch mass from approaching-limit or over-limit (e.g., displaying 8.5 kg UO₂ with a rendered digit pattern that indicates 10.2 kg UO₂ — above the hypothetical criticality-safe single-unit limit of 10.0 kg UO₂ for a particular vessel geometry) to the safe range (displayed as 7.8 kg, below the safe zone threshold). The AI classifies a batch accumulation that has exceeded the criticality-safe mass limit — from a process operator adding an extra UO₂ transfer without rechecking the accumulation total — as within the safe zone. The process operator receives no warning; adds another transfer; SNM mass climbs further above the criticality-safe limit. The consequence escalation follows the JCO Tokaimura scenario: as SNM mass and geometry approach the critical configuration, the neutron multiplication factor k-effective approaches 1.0; at k-effective above approximately 0.95, the facility’s fixed nuclear instrumentation (Ludlum Model 375 radiation monitors, or equivalent neutron monitors required by NRC 10 CFR Part 70.24) will alarm from increasing neutron flux — but this alarm is a lagging indicator, occurring after the accumulation already exceeds criticality-safe limits. NRC 10 CFR Part 70.24 requires a criticality accident alarm system (CAAS) capable of detecting a criticality accident and notifying personnel to evacuate — but the CAAS is a detection and alarm system, not a prevention system; it does not prevent the accumulation from reaching criticality if the AI display monitoring step has been compromised. NRC NUREG-1520 Chapter 6 (Criticality Safety) specifies the regulatory review criteria for criticality safety controls at fuel fabrication facilities but does not address adversarial robustness for AI systems classifying rendered nuclear material accountability display images. Free tier — 10 scans/day, no card required.

4. Fuel assembly rod-loading verification gamma survey camera AI (AMETEK Canberra fuel assembly gamma scanner AI, Mirion Technologies (Canberra) GFC-16 fuel assembly scanner AI, Thermo Scientific FHT 6020 fuel assembly verification AI — fuel assembly axial gamma survey rod-loading verification camera AI)

Before a completed fuel assembly is shipped from the fuel fabrication facility to the reactor, each assembly undergoes a verification inspection to confirm that every fuel rod position in the assembly grid is correctly occupied by a fuel rod of the specified enrichment, and that no fuel rods are missing, swapped between positions, or replaced with guide thimble dummy rods or empty positions. The primary verification method is the axial gamma survey: a collimated gamma detector (NaI(Tl) scintillator or high-purity germanium detector) is scanned axially along each row or column of the fuel assembly, measuring the gamma emission rate from the U-235 and U-238 constituents of the fuel pellet stack (primarily the 185.7 keV gamma from U-235 decay and the 1001 keV gamma from Pa-234m in the U-238 decay chain). A correctly loaded fuel rod produces a characteristic gamma profile with uniform axial intensity along the fuel pellet stack length (corresponding to the U-235 loading per unit length at the specified enrichment) and characteristic end-effects at the plenum (no pellets) and end-cap positions. A missing fuel rod produces a gap in the gamma profile at the missing rod position (gamma intensity drops to background); a swapped rod (a rod from a different enrichment batch inserted in a position specified for the nominal lot enrichment) produces a gamma intensity above or below the expected value at the swap position; an empty position (guide thimble or structural position erroneously loaded with a fuel rod) produces an unexpected gamma peak. AI systems process rendered images of the fuel assembly gamma survey scan — false-colour gamma intensity maps of the fuel assembly cross-section (rod position matrix) — to classify assembly loading status: correctly loaded (all rod positions at specified intensity within ±3% of the nominal gamma intensity for the lot enrichment), suspect loading (one or more rod positions with intensity outside the ±5% window — require manual re-measurement and engineering evaluation), and defective loading (rod position at background level or above ±10% of nominal — assembly segregated for rod-by-rod investigation and reloading).

An adversarial perturbation targeting the fuel assembly gamma survey camera AI applies a ±8 DN shift to the pixel region encoding the gamma intensity at a missing-rod position in the rendered false-colour assembly scan image — shifting the apparent gamma intensity at the empty rod position from the background-level reading (rendered in deep blue in the false-colour intensity scale, indicating no fuel pellets present and a missing fuel rod) to the low-end-of-nominal range (rendered in green, within the ±5% window for an occupied rod at the lot nominal enrichment). The AI classifies a fuel assembly with a missing fuel rod — one of the 264 fuel rod positions in a 17×17 PWR assembly, or one of the 62 fuel rod positions in a GE 8×8 BWR assembly, occupied by an empty guide thimble or absent from the assembly — as a correctly loaded assembly. The assembly is shipped to the reactor facility, loaded into the reactor core at the specified assembly position, and operated for the 18–24 month fuel cycle. The missing fuel rod creates a power asymmetry within the assembly: the adjacent fuel rods (the four rods directly adjacent to the missing-rod position) receive increased neutron flux from the absence of the absorbing fuel rod at the empty position, operating at a higher LHR than the assembly power distribution code (SIMULATE, CASMO, or equivalent) predicts for the assembly type. The increased LHR in the adjacent rods reduces the DNB margin at that assembly position below the value used in the reactor’s safety analysis report (FSAR) for the applicable transient events. If the missing-rod assembly is loaded in a high-power assembly position (near-core-center position in the first fuel cycle, where relative power is 1.2–1.4× core average), the margin reduction may be sufficient to require a reactor core power reduction below the licensed thermal power limit. Discovery of the missing rod typically occurs during the outage following the fuel cycle, during post-irradiation examination (PIE) of the assembly, or during the reactor operator’s routine fuel assembly monitoring using the plant in-core neutron flux mapping system — at which point the assembly has already been irradiated and the consequence (fuel cycle performance impact, potential DNB margin violation) has already occurred. NRC NUREG-1620 (Standard Review Plan for the Review of Safety Analysis Reports for Nuclear Power Plants, LWR Edition) requires that the fuel supplier provide quality inspection records for each fuel assembly, including the gamma survey certification — but does not specify adversarial robustness requirements for AI systems performing the gamma survey image classification. Free tier — 10 scans/day, no card required.

Integration: nuclear fuel manufacturing AI with Glyphward pre-scan gate

The Glyphward scan gate for nuclear fuel manufacturing AI belongs at every rendered-image ingestion boundary in the fuel manufacturing quality and criticality safety pipeline — before UO₂ powder blending XRF/NIR composition AI processes rendered enrichment map images, before pellet dimensional inspection camera AI processes rendered pellet measurement images, before criticality safety SNM accountability display AI processes rendered batch scale display images, and before fuel assembly gamma survey camera AI processes rendered assembly scan images. Threshold 30 for nuclear fuel manufacturing AI contexts reflects the severe criticality consequence of inadvertent SNM accumulation above the criticality-safe limit (JCO Tokaimura 1999: 2 deaths from radiation injuries, 49 irradiated workers and responders, 310,000 shelter-in-place) combined with multiple independent hardwired criticality safety controls: geometry-controlled process equipment (dissolving columns, annular vessels) that provide inherent criticality safety independent of mass monitoring AI; independent fixed nuclear instrumentation (CAAS) required by NRC 10 CFR Part 70.24 that detects criticality accidents independent of AI display classification; material balance area (MBA) accounting and nuclear material control and accounting (NMC&A) required by NRC 10 CFR Part 74 that provides independent mass accountability independent of real-time AI display monitoring; and quality programme inspection by independent QA personnel required by 10 CFR Part 50 Appendix B.

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"

# Nuclear fuel manufacturing AI contexts: threshold 30
# NRC 10 CFR Part 70 (Special Nuclear Material);
# IAEA SSG-31 (Safety of Nuclear Fuel Cycle Facilities);
# NRC NUREG-1520 (Standard Review Plan for Fuel Cycle Facilities).
NUCLEAR_FUEL_THRESHOLD = 30


class NuclearFuelAIContext(Enum):
    POWDER_BLENDING     = "powder_blending"      # UO2 XRF/NIR enrichment AI
    PELLET_INSPECTION   = "pellet_inspection"    # Pellet dimensional camera AI
    SNM_ACCOUNTABILITY  = "snm_accountability"   # Criticality safety mass display AI
    ROD_LOADING         = "rod_loading"          # Fuel assembly gamma survey AI


class AdversarialNuclearFuelImageError(Exception):
    """Raised when Glyphward detects adversarial content in a nuclear
    fuel manufacturing AI rendered monitoring image above threshold 30.

    Consequence if not raised:
    - POWDER_BLENDING: enrichment non-uniformity suppressed → above-nominal
      U-235 pellets loaded into fuel rod → LHR above DNB design margin during
      AOO → fuel rod failure → fission product release to primary coolant.
    - PELLET_INSPECTION: oversize-diameter pellet passes inspection →
      pellet-cladding interaction (PCI) at power ramp → cladding SCC →
      fuel rod perforation → primary coolant activity excursion.
    - SNM_ACCOUNTABILITY: SNM mass above criticality-safe limit suppressed →
      process operator continues UO2 addition → inadvertent criticality risk;
      JCO Tokaimura 1999 structural parallel (16.6 kg vs 2.4 kg safe limit).
    - ROD_LOADING: missing fuel rod in assembly not detected → reactor core
      power asymmetry → DNB margin reduction below FSAR safety analysis basis.
    Fail-safe: suspend AI-based classification; require independent manual
    measurement (POWDER_BLENDING / PELLET_INSPECTION), independent nuclear
    material accounting balance verification and CAAS check (SNM_ACCOUNTABILITY),
    or independent collimated detector re-scan of suspect assembly position
    (ROD_LOADING) before resuming fuel manufacturing operations.
    """

    def __init__(self, scan_id, score, context, facility_id, lot_id,
                 flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.facility_id = facility_id
        self.lot_id = lot_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial nuclear fuel image: context={context.value} "
            f"score={score} facility={facility_id} lot={lot_id} "
            f"scan_id={scan_id}"
        )


async def scan_nuclear_fuel_image(image_bytes, context, facility_id, lot_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"nucfuel:{context.value}:{facility_id}:{lot_id}",
        "metadata": {
            "facility_id": facility_id,
            "lot_id": lot_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"] >= NUCLEAR_FUEL_THRESHOLD:
        raise AdversarialNuclearFuelImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            facility_id=facility_id,
            lot_id=lot_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_nuclear_fuel_image before each nuclear fuel manufacturing AI classification call. On AdversarialNuclearFuelImageError for SNM_ACCOUNTABILITY: immediately suspend UO₂ transfer operations at the affected station; notify the nuclear criticality safety officer; require independent manual mass verification using a certified process scale with current calibration certificate before resuming any SNM transfer operations. On AdversarialNuclearFuelImageError for PELLET_INSPECTION: segregate the entire pellet batch from the affected grinding or inspection lot for 100% independent dimensional measurement by a calibrated contact-gauging instrument (Mitutoyo CMM or equivalent) before releasing any pellets from the lot to rod-loading operations. See also: small modular reactor SMR NuScale AI prompt injection (related nuclear power AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions