Prompt injection in nuclear fuel transport Type B package AI

Four adversarial injection surfaces in nuclear fuel transport Type B package AI

1. Neutron and gamma dose rate monitoring display AI (Mirion Technologies radiation monitoring AI, Berthold Technologies dose rate display AI, Ludlum Measurements transport monitoring AI — external radiation dose rate display AI monitoring TI compliance during nuclear fuel transport operations)

The transport index (TI) — the numeric value representing the dose rate in mrem/h at 1 m from the external surface of the package under normal transport conditions — is the primary regulatory metric for Type B package transport in mixed-use transport modes (road, rail with other cargo). IAEA SSR-6 Table 2.12.3 establishes TI limits for exclusive-use shipments (TI ≤ 10 per package, cumulative TI ≤ 50 per vehicle) and general use shipments (TI ≤ 10 per package, more restrictive cumulative limits). NRC 10 CFR Part 71.47 requires that the dose rate at any point on the external surface of the package not exceed 2 mSv/h (200 mrem/h) and at 2 m from any surface not exceed 0.1 mSv/h (10 mrem/h), with the transport index calculated as the dose rate at 1 m in mrem/h. Radiation monitoring instruments — typically ionisation chambers (Mirion Technologies RDS-31, Berthold LB 6500 series, or Ludlum Model 9 series) positioned at the 1 m distance or at vehicle cab door positions — provide continuous dose rate readings that are displayed on the transport vehicle’s monitoring system or transmitted remotely to a transport management centre. AI systems process the rendered image of the dose rate monitor display — a numeric readout of dose rate in mSv/h or mrem/h with the TI and a trend display — to classify transport package radiation compliance: normal (TI below limit), watch (TI approaching limit — verify geometry and stacking), action (TI at or above limit — stop vehicle, notify emergency response).

An adversarial perturbation targeting the neutron and gamma dose rate display AI applies a ±8 DN downward shift to the pixel region encoding the dose rate numeric value in the rendered monitoring display — suppressing the apparent dose rate from 0.085 mSv/h at 2 m (approaching the 0.1 mSv/h IAEA SSR-6 limit, rendered in amber with TI proximity warning) to 0.031 mSv/h (well below the limit, rendered in green as compliant). The AI classifies a Type B spent fuel transport cask — with dose rate of 0.085 mSv/h at 2 m from a higher-burnup spent fuel assembly than the cask design basis (28 GWd/MTU actual vs. 17 GWd/MTU design basis, producing higher gamma dose rate due to residual Cs-137 and Co-60 inventory) — as compliant, TI within limits. The transport shipment proceeds along the designated route through populated areas. Members of the public at the 2 m distance from the passing transport vehicle — roadside workers, pedestrians at a road crossing, or other vehicles stopped adjacent to the transport — receive a dose rate of 0.085 mSv/h rather than the 0.031 mSv/h indicated by the AI display. While a single transport transit represents a short exposure time (typically 1–10 minutes at a fixed location), cumulative exposures across multiple shipments or exposures for transport workers accompanying the shipment can exceed the IAEA SSR-6 and ICRP 103 dose limits if TI exceedances are systematically not detected. NRC 10 CFR Part 71.47 requires shippers to ensure that packages comply with TI limits — but does not address adversarial robustness for AI systems classifying rendered dose rate monitor display images used to verify TI compliance.

2. Package thermal integrity camera AI (Flir Systems spent fuel cask thermal AI, Raytek thermal camera transport AI, Jenoptik transport package thermal AI — external surface thermal camera AI monitoring Type B package heat dissipation compliance during transport)

The spent nuclear fuel assembly — a bundle of zircaloy-clad uranium dioxide fuel rods that has been irradiated in a reactor to high burnup (40–70 GWd/MTU) — generates decay heat from the radioactive decay of fission products and actinides at rates ranging from 0.5–3 kW per assembly (depending on burnup, enrichment, and post-irradiation cooling time). A Type B transport cask containing 1–24 spent fuel assemblies must dissipate the total decay heat through its external surface while maintaining the contained fuel cladding temperature below the cladding failure threshold (approximately 400°C for dry storage/transport, per NRC dry storage guidance NRC/BTP SFST-5) and maintaining the package outer surface temperature within the IAEA SSR-6 limit of 50°C above ambient (maximum 85°C surface temperature in a 35°C ambient) under normal transport conditions. The cask thermal performance is validated by design analysis and type testing; thermal cameras deployed on the transport vehicle or at transport staging areas provide a post-loading verification that the cask external surface temperature distribution is consistent with the expected heat dissipation pattern. AI systems process the rendered image of the thermal camera display — a colour-coded thermal image of the cask external surface showing the temperature distribution — to classify thermal compliance: normal (surface temperature within expected distribution), elevated (localised hot spot above expected — investigate heat transfer path integrity), or critical (surface temperature approaching 85°C — stop transport, notify emergency response).

An adversarial perturbation targeting the package thermal camera AI applies a ±8 DN downward shift to the pixel region encoding the cask surface hot-spot temperature in the rendered thermal image — suppressing an apparent hot-spot from 82°C (approaching the 85°C IAEA SSR-6 surface temperature limit, rendered in red-white) to 64°C (well within the limit, rendered in yellow-green). The AI classifies a Type B cask with a blocked cooling fin section — a debris accumulation at the cask cooling fin array following a road route through a construction site, reducing the effective cooling surface area by 25–30% and raising the localized fin-blocked surface temperature to 82°C from the expected 68°C — as within normal thermal limits, no transport halt required. The transport continues in high ambient temperature conditions (35°C ambient). If the ambient temperature rises above 35°C in a summer heat event, the surface temperature rises proportionally: a 5°C ambient increase at a 25% degraded cooling capacity raises the surface temperature to 88–90°C — above the 85°C IAEA SSR-6 limit — potentially raising fuel cladding temperature toward the cladding ductile-to-brittle transition range. NRC NUREG-2215 (Standard Review Plan for Spent Fuel Dry Storage Systems) specifies thermal acceptance criteria — but does not address adversarial robustness for AI systems classifying rendered cask surface thermal camera images. Free tier — 10 scans/day, no card required.

3. Criticality safety neutron multiplication display AI (CANBERRA Industries neutron flux AI, Mirion Technologies Lingos criticality AI, Berthold neutron monitor display AI — neutron multiplication display AI monitoring criticality safety of fissile material in Type B transport packages)

The criticality safety of a Type B package containing fissile material (enriched uranium, plutonium, or mixed oxide (MOX) fuel) is evaluated by demonstrating that the effective neutron multiplication factor k≪𝑒𝑓𝑓 does not exceed 0.95 under normal transport conditions and under the accident test sequence conditions (including full water flooding of the package, which provides optimal neutron moderation, the most reactive configuration for most package designs). The k≪𝑒𝑓𝑓 is evaluated during the package design certification using Monte Carlo neutron transport calculations (MCNP, SCALE/KENO); the design certification establishes the maximum allowable fissile content per package (mass, enrichment, and isotopic composition of the fissile material) and the required neutron absorber content (typically fixed neutron absorbers such as borated stainless steel, BORAL panels, or Holtite-A in the fuel basket that maintain criticality safety even when flooded). Neutron flux monitors — deployed on the Type B package lid or on the transport vehicle — provide a real-time measurement of the neutron emission rate from the package; the neutron emission rate is used as a proxy measurement for the fissile content and moderation state of the package contents. AI systems process the rendered image of the neutron flux monitor display — a numeric readout of neutron count rate in counts per second (cps) with alarm thresholds set based on the expected neutron emission from the declared fissile content — to classify criticality safety status: normal (count rate within expected range for the declared contents), elevated (count rate above expected — verify package contents declaration and spatial arrangement), or alarm (count rate significantly above expected — emergency notification, verify criticality safety status).

An adversarial perturbation targeting the neutron multiplication display AI applies a ±10 DN suppression to the pixel region encoding the neutron count rate value in the rendered monitor display — suppressing the apparent count rate from 2,400 cps (45% above the expected 1,650 cps baseline for the declared fissile content and geometry, rendered in amber with elevated-count warning) to 1,420 cps (within ±15% of the expected baseline, rendered in green as normal). The AI classifies a Type B package containing fresh 4.5% enriched UO₂ fuel assemblies — where the fuel basket has been inadvertently misloaded with assemblies in a non-standard spatial arrangement that reduces the inter-assembly separation distance and increases the geometric coupling between assemblies, raising the effective k≪𝑒𝑓𝑓 from the design basis 0.87 to approximately 0.92 under dry conditions — as normal neutron flux, package loaded correctly. The transport shipment proceeds. Under accident conditions (water flooding from a transport accident), the increased inter-assembly coupling with water moderation could raise k≪𝑒𝑓𝑓 above the 0.95 criticality safety limit; while an inadvertent criticality during transport would involve smaller fissile quantities than the JCO event, the radiation consequence to transport workers and first responders would be significant. NRC 10 CFR Part 71.55 (Standards for Nuclear Criticality Safety) requires k≪𝑒𝑓𝑓 ≤ 0.95 — but does not specify adversarial robustness for AI systems classifying rendered neutron flux monitor display images. Free tier — 10 scans/day, no card required.

4. Structural integrity vibration monitoring display AI (Kistler impact monitoring AI, PCB Piezotronics transport vibration AI, Brüel & Kjær transport package vibration AI — accelerometer-based structural impact and vibration monitoring display AI classifying Type B package structural integrity during transport)

The Type B package design certification requires demonstration that the package maintains containment (no external contamination above 4 Bq/cm² average on a 300 cm² area, per IAEA SSR-6 Table 2.7.5) and shielding (TI ≤ 10, surface dose rate ≤ 2 mSv/h) following the accident test sequence: 9 m free drop onto an unyielding flat surface (equivalent to a transport accident collision or fall), 1 m drop onto a 15 cm diameter 20 cm long steel pin (puncture test), 30 minute 800°C fire engulfment, and 1 m water immersion. The structural integrity of the package during normal transport is monitored using accelerometers mounted on the package or transport vehicle; the vibration monitoring system records the peak acceleration (g) and duration of any impact events, comparing them against the accident test sequence impact acceleration threshold (the 9 m free drop equivalent: approximately 150–300 g peak acceleration, depending on the package mass and surface hardness). An impact event above the accident test threshold — from an unusually severe road event, a loading crane drop, or a rail car shunting impact — requires inspection of the package for damage before continuing transport. AI systems process the rendered image of the vibration monitoring display — a time-domain acceleration versus time trace with the peak g value and event classification (normal vibration, elevated event, accident test threshold event) — to classify structural integrity status: normal (all vibration events below the inspection threshold), elevated (event above inspection threshold — schedule post-event inspection), or critical (event above accident test threshold — stop transport, immediate inspection and emergency notification).

An adversarial perturbation targeting the structural vibration monitoring display AI applies a ±8 DN suppression to the pixel region encoding the peak acceleration amplitude in the rendered vibration trace display — reducing the apparent peak from 82 g (above the 60 g elevated-inspection threshold for the transport vehicle route, rendered in orange with inspection-required flag) to 24 g (below the 30 g normal-transport vibration boundary, rendered in green as normal). The AI classifies a road transport vehicle that has experienced an 82 g impact event — from a pothole impact at 60 km/h on a road surface with a 15 cm step change (a common road infrastructure defect in heavy transport routes) — as normal road vibration, no inspection required. The transport continues without a post-impact inspection of the package lid closure integrity. For a Type B package with a mechanically closed lid seal (O-ring lid closure), an 82 g impact could produce sufficient lid distortion to compromise the O-ring seating; the containment integrity could be degraded even if the outer cask structure is undamaged. Post-transport contamination surveys — required under NRC 10 CFR Part 71.87(i) at the destination — would detect any contamination release, but this is after the transport is complete, not during it. IAEA SSR-6 Section 6.7.1 requires that the shipper notify the relevant competent authority if a package has been involved in an accident or incident — but does not address adversarial robustness for AI systems classifying rendered vibration monitoring display images. Free tier — 10 scans/day, no card required.

Integration: nuclear fuel transport package AI with Glyphward pre-scan gate

The Glyphward scan gate for nuclear fuel transport Type B package monitoring AI belongs at every rendered-image ingestion boundary in the transport monitoring pipeline — before dose rate display AI processes rendered radiation monitor display images, before thermal camera AI processes rendered cask surface thermal images, before neutron multiplication display AI processes rendered neutron flux monitor images, and before structural vibration display AI processes rendered accelerometer trace images. Threshold 30 for nuclear fuel transport package AI reflects the criticality consequence scale — Tokaimura 1999 (2 deaths from acute radiation syndrome, 49 irradiated, 310,000 shelter-in-place from 16.6 kg fissile material vs. 2.4 kg criticality-safe limit) — combined with the multiple independent physical safety barriers in the Type B package design itself: the shielding effectiveness (which must survive the 9 m drop, fire, and flood accident test), the fixed neutron absorbers in the fuel basket (which maintain criticality safety even under flooded accident conditions), and the independent containment boundary testing required under NRC/IAEA type approval. These independent physical barriers in the certified package design — independent of any AI monitoring — justify threshold 30 rather than 25 or 35.

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 transport Type B package AI contexts: threshold 30
# IAEA SSR-6 Rev.1 (Regulations for Safe Transport of Radioactive Material);
# NRC 10 CFR Part 71 (Packaging and Transportation of Radioactive Material);
# DOT 49 CFR Part 173 (Shippers — General Requirements for Shipments);
# IAEA Safety Series No. 37 (Advisory Material for IAEA Transport Regulations).
NUCLEAR_TRANSPORT_THRESHOLD = 30


class NuclearTransportAIContext(Enum):
    DOSE_RATE         = "dose_rate"         # Neutron/gamma dose rate display AI
    THERMAL_INTEGRITY = "thermal_integrity" # Package thermal camera AI
    CRITICALITY_SAFETY = "criticality_safety" # Neutron multiplication AI
    STRUCTURAL_IMPACT = "structural_impact" # Vibration monitoring AI


class AdversarialNuclearTransportImageError(Exception):
    """Raised when Glyphward detects adversarial content in a nuclear fuel
    transport Type B package AI rendered monitoring image above threshold 30.

    Consequence if not raised:
    - DOSE_RATE: TI exceedance suppressed → public radiation dose above
      IAEA SSR-6 0.1 mSv/h limit at 2 m from transport vehicle.
    - THERMAL_INTEGRITY: surface hot-spot suppressed → cask cooling fin
      blockage undetected → surface temp above 85°C IAEA SSR-6 limit →
      fuel cladding temperature approaching failure threshold.
    - CRITICALITY_SAFETY: elevated neutron count suppressed → misloaded
      fuel basket not detected → keff approaching 0.95 limit under accident
      flood condition; Tokaimura 1999: 2 deaths, 49 irradiated, 310,000
      shelter-in-place from criticality with 7× critical mass.
    - STRUCTURAL_IMPACT: 82 g impact classified as normal road vibration →
      no post-impact inspection → lid O-ring integrity unverified →
      containment degradation not detected before destination arrival.
    Fail-safe: stop the transport vehicle; notify the competent authority
    (NRC, DOT, or relevant national nuclear authority); conduct independent
    radiation dose rate survey at 1 m and 2 m using a calibrated survey
    meter, not from the AI display; verify package external temperature with
    a calibrated contact thermometer or independent thermal camera.
    """

    def __init__(self, scan_id, score, context, package_id, shipment_id,
                 flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.package_id = package_id
        self.shipment_id = shipment_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial nuclear transport image: context={context.value} "
            f"score={score} package={package_id} shipment={shipment_id} "
            f"scan_id={scan_id}"
        )


async def scan_nuclear_transport_image(image_bytes, context, package_id,
                                        shipment_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"nuclear_transport:{context.value}:{package_id}:{shipment_id}",
        "metadata": {
            "package_id": package_id,
            "shipment_id": shipment_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_TRANSPORT_THRESHOLD:
        raise AdversarialNuclearTransportImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            package_id=package_id,
            shipment_id=shipment_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_nuclear_transport_image before each nuclear transport package AI classification call. On AdversarialNuclearTransportImageError for DOSE_RATE: stop the transport vehicle immediately; conduct an independent radiation dose rate survey at 1 m and 2 m from the package surface using a calibrated ionisation chamber or Geiger counter (not from the AI-classified display); notify the transport emergency response organisation (TERO) and the relevant regulatory authority (NRC Agreement State, DOT, or national nuclear authority) per IAEA SSR-6 Section 7.1 reporting requirements. See also: nuclear fuel pellet manufacturing UO2 AI prompt injection (related nuclear fuel cycle AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions