Prompt injection in insurance underwriting and actuarial AI

Four adversarial injection surfaces in insurance underwriting and actuarial AI

1. Property aerial and satellite inspection photograph injection (EagleView AI, CoreLogic AI, Verisk/ISO underwriting AI)

Property aerial and satellite inspection photograph AI processes aerial images captured by EagleView’s fixed-wing aircraft and satellite constellation, CoreLogic’s property data imagery platform, and third-party aerial survey providers submitted through Verisk ISO underwriting AI property risk scoring tools, regional carrier underwriting workstation interfaces, and surplus lines eligibility assessment platforms that extract roof condition grade values, hail damage classification flags, non-standard construction type indicators, wind mitigation feature identifications, and property hazard proximity risk scores from aerial inspection image inputs, generating ISO rating bureau risk class assignments, territory classification codes, and underwriting eligibility determinations that govern whether a property qualifies for standard admitted market coverage at the state DOI-filed rate plan, requires surplus lines market placement, or is declined entirely on the basis of AI-assessed physical condition and hazard exposure. EagleView AI has analysed more than 700 million properties using aerial imagery and machine learning, with its roof condition assessment outputs routinely incorporated into Verisk carrier client underwriting workflows as a primary or supplementary data source for roof age, material type, condition grade, and hail impact damage determination — outputs that directly affect whether a carrier applies a roof cosmetic damage exclusion endorsement, assigns a roof-age premium surcharge, or declines coverage on the basis of roof condition at policy inception or renewal. CoreLogic AI processes more than one billion property records through AI-assisted property risk analytics, location intelligence, and hazard exposure assessment tools that primary carriers and mortgage lenders requiring force-placed insurance depend upon for replacement cost estimation, flood zone classification, wildfire risk scoring, and property condition assessment — functions that aggregate into the property-level risk profile that informs admitted market eligibility and E&S market referral decisions across the US residential and commercial property insurance market.

The adversarial injection surface is the aerial or satellite inspection photograph submission pathway: EagleView or CoreLogic aerial property inspection images submitted through Verisk ISO underwriting AI property risk scoring interfaces or regional carrier underwriting workstations for AI roof condition assessment, hail damage classification, construction type identification, and hazard proximity scoring. An adversarially crafted aerial property inspection photograph — in which pixel perturbations applied to the roof surface shingle deterioration indicator, hail impact dimple pattern visual marker, or non-standard construction material identification region on an EagleView aerial image cause the AI to classify the roof as a standard-condition shingled roof meeting preferred underwriting criteria when the actual image documents a deteriorated, aged, or hail-damaged roof meeting the carrier’s adverse underwriting conditions for cosmetic exclusion, rated surcharge, or eligibility decline — can suppress a roof condition deficiency classification that would otherwise trigger a premium surcharge, coverage limitation endorsement, or eligibility denial determination in the Verisk underwriting AI workflow. In high-volume residential property insurance renewal processing environments where EagleView aerial inspection photographs are ingested through automated Verisk underwriting AI workflows for thousands of properties per day during catastrophe-impacted renewal cycles, adversarial suppression of a hail damage classification across a cohort of affected properties allows systematically underpriced hail-loss risk to persist in the carrier’s book of business, creating adverse loss ratio deterioration that may not surface until the catastrophe year combined ratio is calculated months after the affected renewal cohort was processed through the adversarially manipulated AI workflow.

The regulatory consequences of adversarially suppressed roof condition deficiency detection in property aerial inspection AI span state DOI rate filing compliance and NAIC model law dimensions of substantial severity. ISO/Verisk P&C underwriting rate and class plans are filed with state Departments of Insurance under state insurance code rate filing requirements and NAIC Property/Casualty Model Laws; adversarial AI manipulation that causes properties with documented hail damage or roof deterioration to be rated in standard-condition ISO risk classes at standard-tier rates — when the filed rate plan specifies surcharges or adverse class assignment for those conditions — creates a rate filing compliance failure in which filed rates are not being applied as filed, with state DOI market conduct examination exposure and potential premium deficiency order consequences. NAIC Model Regulation 900 (Unfair Claims Settlement Practices) applies to the claims settlement dimension of AI-suppressed property deficiency: when a carrier settles a property loss claim arising from a condition that was present at underwriting but suppressed by adversarially manipulated AI, the claims settlement may become subject to NAIC Model 900 scrutiny on the grounds that the carrier’s AI underwriting workflow failed to identify and disclose a material property condition affecting coverage scope. State DOI property insurance rate and form examination cycles assess whether carrier AI tools produce outcomes consistent with filed rate and form plans; adversarial manipulation that systematically misclassifies a cohort of properties in adverse condition as standard-tier will appear as an unexplained loss ratio anomaly on DOI examination, but the examination process does not inspect individual aerial inspection image pixels for adversarial manipulation — it observes statistical outcomes rather than the per-image AI input boundary. Threshold: 55 for property aerial inspection AI — reflecting the premium adequacy, rate filing compliance, and adverse selection dimensions of suppressed property condition risk classification.

2. Natural catastrophe model loss visualisation injection (RMS AI, Swiss Re AI, Munich Re AI)

Natural catastrophe model loss visualisation AI processes screenshots of RMS One platform probabilistic loss scenario distribution displays, exceedance probability curve output images, Swiss Re CatNet aggregate exposure accumulation map visualisations, Munich Re NATHAN Risk Suite loss scenario display screenshots, and reinsurance treaty underwriting submission catastrophe model summary report images submitted through AI-assisted reinsurance treaty underwriting, retrocession risk pricing, and catastrophe risk quantification tools that extract probable maximum loss (PML) point estimates, tail value at risk (TVaR) classifications, annual aggregate exceedance probability curve shape assessments, and geographic exposure accumulation anomaly flags from catastrophe model output image inputs, generating treaty pricing recommendations, retrocession programme adequacy evaluations, and NAIC Risk-Based Capital formula catastrophe reserve loading determinations that govern whether a reinsurance treaty is written at a given limit and premium, whether a retrocession programme provides adequate protection for the ceded portfolio’s tail risk, and whether an insurer or reinsurer’s catastrophe risk capital reserve meets NAIC RBC formula requirements or Solvency II Solvency Capital Requirement (SCR) standards. RMS AI operates the RMS One catastrophe modelling platform — the industry’s most widely deployed probabilistic catastrophe loss quantification infrastructure — across more than 100 territories and more than 50 perils including US hurricane, US earthquake, European windstorm, Japan typhoon, and global flood, with RMS One output visualisations processed by AI-assisted tools at primary insurers, reinsurers, Lloyd’s syndicates, and catastrophe bond issuers whose capital adequacy and treaty underwriting decisions depend on the fidelity of AI-generated catastrophe loss interpretations to the underlying model output data. Swiss Re AI processes global reinsurance treaty submissions including catastrophe model output screenshot images through AI-assisted underwriting tools at Swiss Re’s property and casualty reinsurance underwriting operations, where AI-generated catastrophe loss interpretations inform pricing and capacity allocation decisions at premium volumes that make per-treaty AI errors consequential at global portfolio scale. Munich Re AI processes Lloyd’s of London market syndicate submissions and specialty reinsurance treaty documents, incorporating catastrophe model loss scenario visualisation screenshots through AI-assisted underwriting tools deployed across Lloyd’s syndicates and the London market specialty lines market.

The adversarial injection surface is the catastrophe model loss visualisation screenshot submission pathway: RMS One platform output screenshots, Swiss Re CatNet exposure accumulation map images, and Munich Re NATHAN Risk Suite loss scenario display images submitted through AI-assisted reinsurance treaty underwriting interfaces or retrocession pricing platforms for AI PML estimation, exceedance probability classification, and aggregate exposure accumulation assessment. An adversarially crafted RMS One exceedance probability curve screenshot — in which pixel perturbations applied to the 1-in-250-year return period PML indicator, the 1-in-100-year TVaR display, or the aggregate accumulation threshold alert region on an RMS One platform output screenshot cause the AI to classify the portfolio’s catastrophe risk profile as within standard reinsurance programme adequacy parameters when the actual screenshot displays a PML exceedance or aggregate accumulation anomaly that would require additional reinsurance layer purchase, retrocession treaty modification, or NAIC RBC catastrophe reserve loading adjustment — can suppress a PML exceedance indicator that would otherwise trigger a capital reserve loading determination, allowing a reinsurance treaty to be written at inadequate premium for the true catastrophe tail risk or allowing an insurer to file a NAIC RBC return that underreports catastrophe reserve loading requirements. In reinsurance treaty renewal cycles where AI-assisted tools process multiple catastrophe model output visualisations from cedants’ submitted RMS One reports, adversarial suppression of an aggregate accumulation anomaly flag across a portfolio of treaty submissions concentrates understated catastrophe risk in the reinsurer’s accepted book, creating portfolio-level capital adequacy exposure that surfaces only when a catastrophe event produces losses that exceed the reinsurer’s NAIC RBC-loaded catastrophe reserve position.

The regulatory consequences of adversarially suppressed catastrophe model PML exceedance detection in reinsurance underwriting AI span NAIC RBC capital formula, Lloyd’s UASG, Solvency II, and reinsurance contract warranty dimensions of exceptional regulatory severity. The NAIC Risk-Based Capital formula for property and casualty insurers and reinsurers includes a catastrophe reserve loading factor derived from probable maximum loss estimates at specified return periods; adversarial AI manipulation that suppresses a PML exceedance indicator in catastrophe model output AI processing causes the insurer or reinsurer to understate the catastrophe loading in its NAIC RBC formula calculation, creating a capital adequacy deficiency that exposes the entity to NAIC RBC company action level or regulatory action level consequences when the understatement is detected in state DOI RBC examination. Lloyd’s of London Underwriting Agency Standards and Guidance (UASG) specifies catastrophe model usage standards and aggregate accumulation monitoring requirements for Lloyd’s syndicates; adversarial manipulation of catastrophe model visualisation AI tools used in Lloyd’s syndicate underwriting creates a UASG compliance failure with Lloyd’s Franchise Board oversight consequences and potential syndicate capacity restriction. Solvency II (EU Directive 2009/138/EC) requires insurers and reinsurers subject to EU regulation to calculate Solvency Capital Requirement (SCR) using approved internal models or the Solvency II standard formula, with catastrophe risk being a material SCR component; adversarial suppression of a PML exceedance in catastrophe model AI output that causes an SCR understatement creates a Solvency II capital adequacy deficiency with national supervisory authority enforcement consequences under Article 137 SCR non-compliance provisions. Reinsurance contract warranty and representation clauses in treaty reinsurance agreements typically include cedant representations regarding the accuracy and completeness of catastrophe model submissions; adversarial manipulation of catastrophe model visualisation screenshots submitted to AI-assisted reinsurance underwriting tools may create breach-of-warranty exposure for the cedant if the manipulation affects the accuracy of catastrophe model representations made in the treaty underwriting submission. Threshold: 60 for catastrophe model visualisation AI — reflecting the capital adequacy, solvency regulatory, and reinsurance contract warranty dimensions of suppressed PML exceedance detection.

3. Auto telematics and dashcam photograph injection (Clearcover AI, Lemonade AI UBI)

Auto telematics and dashcam photograph AI processes dashcam video frame images captured by policyholder-installed dashcam devices, OBD-II telematics display screenshots from connected vehicle diagnostic interfaces, and driver behaviour summary visualisation images generated by usage-based insurance (UBI) platform data aggregation tools submitted through Clearcover AI usage-based premium rating workflows, Lemonade AI auto insurance telematics assessment interfaces, and UBI platform AI risk scoring tools that extract distracted driving behaviour anomaly classifications, speeding pattern frequency scores, hard-braking event count grades, nighttime driving proportion risk indicators, and phone-use-while-driving detection flags from telematics image inputs, generating UBI driver risk tier assignments, per-renewal premium surcharge or discount calculations, and policy non-renewal eligibility flags that govern whether a UBI policyholder receives a safe-driver discount, a distracted-driving surcharge, or a non-renewal determination at the UBI policy renewal cycle under the carrier’s state DOI-filed UBI rate plan. Clearcover AI deploys usage-based insurance rating through a direct-to-consumer digital auto insurance platform that integrates telematics data processing and driver behaviour AI scoring as a core premium rating input, with dashcam photograph frame analysis and OBD telematics display image processing generating the driver behaviour risk scores that determine individual policyholder UBI premium tier placement at renewal. Lemonade AI processes auto insurance telematics data through AI-assisted UBI premium rating workflows that incorporate driver behaviour visualisation images and telematics display screenshots as inputs to AI risk classification tools, generating UBI premium rate adjustments that Lemonade applies to policyholder renewal premiums under state DOI-filed UBI rate plans. The broader UBI market context in which Clearcover and Lemonade operate is governed by NAIC Automated Vehicles Insurance Working Group (AVIWG) guidance on AI-assisted vehicle risk classification, state DOI auto insurance rate filing requirements for UBI programmes, and NAIC UBI Model Law provisions specifying data collection, usage, and disclosure requirements for telematics-based auto insurance rating — a regulatory framework that creates compliance exposure when AI-assisted UBI rating tools produce outcomes inconsistent with the disclosed UBI scoring methodology.

The adversarial injection surface is the dashcam photograph frame and OBD telematics display image submission pathway: dashcam video frames and telematics display screenshots submitted through Clearcover AI or Lemonade AI UBI premium rating interfaces for AI driver behaviour anomaly classification, speeding pattern detection, hard-braking event frequency scoring, and nighttime driving risk assessment. An adversarially crafted dashcam photograph frame — in which pixel perturbations applied to the driver eye gaze region, steering wheel grip indicator, or mobile phone screen reflection on a dashcam forward-facing or cabin-facing camera frame cause the AI to classify a phone-use-while-driving event or distracted driving behaviour as attentive standard driving when the actual image documents a distracted driving event meeting the carrier’s UBI rate plan criteria for a distracted-driving anomaly flag and associated premium surcharge — can suppress a distracted driving anomaly classification that would otherwise generate a UBI premium surcharge at the policyholder’s next renewal cycle, allowing a high-risk UBI driver to be rated at a lower-risk premium tier than the carrier’s filed UBI rate plan specifies for that driver’s documented behaviour pattern. In digital-first UBI auto insurance programmes where AI-assisted telematics image processing handles hundreds of thousands of dashcam frame assessments per renewal cycle without individual human adjuster review, adversarial suppression of distracted driving anomaly flags across a cohort of high-risk UBI policyholders degrades the actuarial accuracy of the UBI rate plan’s loss cost predictions, producing adverse loss ratios that the carrier’s UBI pricing actuaries cannot attribute to identifiable rating factor anomalies because the distorted AI inputs are invisible to the loss cost analysis that the actuaries perform on AI-output data rather than on source images.

The regulatory consequences of adversarially suppressed driver behaviour anomaly detection in auto telematics AI span state DOI auto insurance rate filing, NAIC AVIWG guidance, and state UBI disclosure requirement dimensions. State DOI auto insurance rate filing requirements mandate that UBI carriers file their telematics scoring methodology, data elements, and rating factor algorithms with the state insurance department; adversarial AI manipulation that systematically produces driver behaviour scores inconsistent with the disclosed scoring methodology — because adversarially crafted dashcam frames cause the AI to output scores at odds with the driving behaviour actually captured in the dashcam image — creates a rate filing compliance failure in which the AI-generated UBI scores are not produced by the disclosed scoring methodology as filed, with state DOI market conduct examination exposure and potential premium refund order consequences. NAIC Automated Vehicles Insurance Working Group guidance specifies that carriers deploying AI-assisted vehicle risk classification tools should implement validation controls to verify that AI output is consistent with underlying telematics data inputs; adversarial dashcam image manipulation that produces systematically incorrect AI driver risk classifications creates an AVIWG guidance conformance failure at the AI validation control level. NAIC UBI Model Law provisions require carriers to provide policyholders with access to the telematics data used in their UBI premium calculation; a policyholder whose dashcam frames were adversarially manipulated to suppress distracted driving anomaly flags and then queries the carrier for their telematics scoring data will receive UBI scoring data that reflects adversarially manipulated AI output rather than their actual driving behaviour record, creating a data accuracy and disclosure compliance dimension to the adversarial injection exposure. Threshold: 60 for auto telematics and dashcam AI — reflecting the UBI rate filing compliance, actuarial loss cost accuracy, and NAIC AVIWG guidance conformance dimensions of suppressed driver behaviour risk classification.

4. Claims damage assessment photograph injection (Lemonade AI, ISO/Verisk ClaimSearch AI)

Claims damage assessment photograph AI processes property damage and auto damage claim photographs submitted by policyholders, claimant-appointed public adjusters, and independent adjusters through Lemonade AI’s AI-assisted claims adjudication workflow, ISO/Verisk ClaimSearch AI’s fraud detection platform, and regional carrier AI claims assessment tools that extract damage type classifications, coverage applicability determinations, loss severity estimates, fraud indicator pattern scores, staged-loss scenario probability values, prior damage identification flags, and inflated estimate anomaly alerts from claims image inputs, generating claims payment authorisation decisions, special investigation unit (SIU) referral flags, claim denial determinations, and settlement amount calculations that govern whether a property or auto damage claim is paid, investigated for fraud, or denied at the initial adjudication stage without human adjuster review. Lemonade AI processes renters and homeowners insurance claims through an AI-assisted adjudication workflow that evaluates policyholder-submitted property damage photographs and supporting evidence images, achieving claimed 30-second claims approval cycles for categorised straightforward loss events by enabling AI classification of damage type, coverage applicability, and settlement amount to proceed without routing to a human adjuster; this speed advantage — which drives Lemonade’s claims experience differentiation and customer satisfaction positioning — depends entirely on the fidelity of AI damage assessment to the actual content of the submitted claim photographs. ISO/Verisk ClaimSearch AI processes a fraud detection database exceeding one billion claims records, with AI-assisted pattern matching against prior claims, staged-loss scenario libraries, and inflated damage indicator classifiers that identify fraud-indicator-bearing claim photographs for SIU referral; ClaimSearch AI is deployed across the majority of US P&C carriers, making its fraud classification outputs consequential to the industry’s aggregate fraud loss experience. Hippo Insurance AI processes home monitoring sensor data and property condition images through AI-assisted claims adjudication support tools that assess whether damage reported in a claim photograph is consistent with sensor data indicating a loss event at the reported time and location, providing a cross-validation data source that informs Hippo’s AI claims adjudication workflow.

The adversarial injection surface is the property damage and auto damage claim photograph submission pathway: policyholder-submitted or adjuster-submitted claim photographs ingested through Lemonade AI claims adjudication interfaces or ISO/Verisk ClaimSearch AI fraud detection platforms for AI damage classification, fraud indicator pattern scoring, prior damage identification, and settlement amount determination. An adversarially crafted property damage claim photograph — in which pixel perturbations applied to the pre-existing wear indicator region, the staged-loss staging artefact visual marker, or the inflated severity anchor in a water damage or fire damage claim photograph cause the AI to classify a claim with multiple fraud indicator patterns as a straightforward legitimate loss meeting Lemonade AI’s 30-second approval criteria when the actual image documents damage characteristics meeting ClaimSearch AI’s staged-loss scenario detection thresholds for SIU referral — can suppress a fraud indicator classification that would otherwise route the claim to human adjuster review and SIU investigation, allowing a fraudulent or materially inflated claim to receive AI-authorised payment without the review that the carrier’s claims handling standards and state DOI unfair claims settlement practice requirements mandate for fraud-indicator-bearing claims. In digital-first insurer claims workflows where AI-assisted adjudication processes thousands of claim photographs per day without human adjuster review for claims below AI-determined severity thresholds, adversarial suppression of fraud indicator classifications enables systematic claims leakage at volumes that may not surface as a statistical anomaly in the carrier’s quarterly fraud analytics review until the cumulative inflated payment total exceeds the threshold for actuarial reserve adequacy review.

The regulatory and criminal consequences of adversarially suppressed claims fraud detection span state DOI insurance fraud statutes, 18 USC §1033, NAIC Model 750, and False Claims Act dimensions of exceptional severity. State DOI insurance fraud statutes in all 50 states impose civil and criminal liability for insurance fraud including staged-loss schemes, inflated damage claims, and material misrepresentation in the claims process; adversarial manipulation of claims damage assessment AI that suppresses fraud indicator classifications and enables fraudulent claims to receive AI-authorised payment may constitute participation in the underlying insurance fraud scheme under state fraud statute accessory or facilitation theories, with civil penalty and criminal exposure for the actor who crafted the adversarial claim photographs. 18 USC §1033 (Crimes by or affecting persons engaged in the business of insurance) imposes federal criminal liability for knowing and wilful commission of a fraudulent act affecting an insurer whose activities affect interstate commerce, with felony penalties including imprisonment up to 10 years; adversarial manipulation of Lemonade AI or ClaimSearch AI claims fraud detection tools that causes fraudulent claims to receive payment constitutes a 18 USC §1033 predicate act with federal criminal exposure at the individual and organisational levels. NAIC Model Regulation 750 (Insurance Fraud Prevention Model Act) requires carriers to implement fraud prevention plans including special investigation unit programmes, fraud reporting procedures, and anti-fraud training; adversarial manipulation of AI-assisted fraud detection tools that defeats a carrier’s Model 750-compliant fraud prevention programme creates a Model 750 compliance failure with state DOI market conduct examination consequences. False Claims Act (31 USC §§3729-3733) creates federal civil liability for fraudulent claims submitted to federal programmes including federal flood insurance, federal crop insurance, and Medicare supplemental insurance; adversarial suppression of AI fraud detection in claims affecting federal programme payors creates False Claims Act qui tam exposure with treble damages and civil penalty consequences. Threshold: 60 for claims damage assessment AI — reflecting the insurance fraud criminal liability, state DOI unfair claims settlement practice, and NAIC Model 750 compliance dimensions of suppressed fraud indicator classification.

Integration: insurance underwriting and actuarial AI image ingestion with Glyphward pre-scan

Insurance underwriting and actuarial AI image ingestion flows from EagleView and CoreLogic aerial property inspection photograph APIs, RMS One and Swiss Re/Munich Re catastrophe model visualisation screenshot channels, Clearcover and Lemonade UBI telematics dashcam photograph interfaces, and Lemonade AI and ClaimSearch claims damage assessment photograph portals into property risk scoring AI, catastrophe model interpretation AI, driver behaviour risk AI, and claims fraud detection AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary before AI-generated output is committed to ISO rating bureau risk class assignments, reinsurance treaty catastrophe reserve determinations, UBI driver risk tier placements, or claims payment authorisation records:

import asyncio
import base64
import hashlib
import os
import uuid
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = os.environ["GLYPHWARD_API_KEY"]
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Insurance underwriting & actuarial AI — state DOI rate filing; NAIC Model
# 900/750; NAIC RBC formula; Lloyd’s UASG; Solvency II; 18 USC §1033.
# Suppression of roof deficiency flags, PML exceedances, driver behaviour
# anomalies, and claims fraud patterns create premium adequacy, capital reserve,
# and insurance fraud regulatory and criminal liability consequences.
THRESHOLD_PROPERTY_AERIAL = 55  # EagleView/CoreLogic/Verisk; ISO rate filing
THRESHOLD_INSURANCE_AI    = 60  # cat model, telematics/UBI, claims fraud


class InsuranceAIContext(str, Enum):
    PROPERTY_AERIAL    = "property_aerial"     # EagleView, CoreLogic, Verisk/ISO
    CAT_MODEL_DISPLAY  = "cat_model_display"   # RMS One, Swiss Re, Munich Re
    TELEMATICS_DASHCAM = "telematics_dashcam"  # Clearcover, Lemonade UBI
    CLAIMS_DAMAGE      = "claims_damage"       # Lemonade AI, ClaimSearch AI


def threshold_for(context: InsuranceAIContext) -> int:
    if context == InsuranceAIContext.PROPERTY_AERIAL:
        return THRESHOLD_PROPERTY_AERIAL
    return THRESHOLD_INSURANCE_AI


async def scan_insurance_ai_image(
    image_path: str | Path,
    context: InsuranceAIContext,
    carrier_id_hash: str,   # SHA-256 of carrier NAIC company code
    policy_ref: str,        # e.g. "POL-HO3-2026-448821", "AUTO-UBI-8847723"
    inspection_id: str,     # EagleView report ID, RMS run ID, dashcam session, claim no.
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan an insurance underwriting or actuarial AI image for adversarial
    injection payloads before forwarding to property risk scoring, catastrophe
    model interpretation, UBI driver behaviour rating, or claims fraud
    detection AI systems.

    Raises AdversarialInsuranceAIImageError if score meets threshold:
      - PROPERTY_AERIAL:    threshold 55; ISO/Verisk rate filing; NAIC Model 900;
                            NAIC P/C Model Laws; state DOI market conduct
      - CAT_MODEL_DISPLAY:  threshold 60; NAIC RBC formula; Lloyd’s UASG;
                            Solvency II SCR; reinsurance contract warranty
      - TELEMATICS_DASHCAM: threshold 60; state DOI UBI rate filing; NAIC AVIWG;
                            NAIC UBI Model Law; state UBI disclosure requirements
      - CLAIMS_DAMAGE:      threshold 60; 18 USC §1033; NAIC Model 750;
                            state DOI fraud statutes; False Claims Act
    """
    image_bytes  = Path(image_path).read_bytes()
    image_b64    = base64.b64encode(image_bytes).decode()
    image_sha256 = hashlib.sha256(image_bytes).hexdigest()
    client_scan_id = str(uuid.uuid4())
    threshold = threshold_for(context)

    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json={
            "image": image_b64,
            "source": context.value,
            "metadata": {
                "insurance_context": context.value,
                "carrier_id_hash":   carrier_id_hash,
                "policy_ref":        policy_ref,
                "inspection_id":     inspection_id,
                "client_scan_id":    client_scan_id,
                "image_sha256":      image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "carrier_id_hash":   carrier_id_hash,
        "policy_ref":        policy_ref,
        "inspection_id":     inspection_id,
        "insurance_context": context.value,
        "scan_id":           result["scan_id"],
        "client_scan_id":    client_scan_id,
        "image_sha256":      image_sha256,
        "score":             result["score"],
        "flagged_region":    result.get("flagged_region"),
        "threshold":         threshold,
        "action":            "blocked" if result["score"] >= threshold else "allowed",
    }
    await write_insurance_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialInsuranceAIImageError(
            f"Insurance AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"carrier={carrier_id_hash} policy={policy_ref}"
        )
    return result


async def write_insurance_audit_record(record: dict) -> None:
    """Persist audit record to carrier compliance audit store (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialInsuranceAIImageError(Exception):
    """Raised when an insurance underwriting or actuarial AI image exceeds the adversarial injection threshold."""
    pass

Call scan_insurance_ai_image() with InsuranceAIContext.PROPERTY_AERIAL before forwarding EagleView or CoreLogic aerial inspection photographs to Verisk ISO underwriting AI property risk scoring tools — the integration point where adversarial suppression of a roof condition deficiency creates a rate filing compliance failure, with inspection_id set to the EagleView report identifier linking the Glyphward scan record to the specific aerial inspection event for state DOI market conduct examination audit trail purposes. Call with InsuranceAIContext.CAT_MODEL_DISPLAY for RMS One output screenshots or Swiss Re/Munich Re catastrophe model visualisation images before AI PML estimation and exceedance probability classification, preserving image_sha256 as the forensic anchor for NAIC RBC formula catastrophe reserve loading audit and Lloyd’s UASG aggregate accumulation monitoring compliance documentation. Call with InsuranceAIContext.TELEMATICS_DASHCAM for Clearcover or Lemonade dashcam photograph frames and OBD telematics display images before AI driver behaviour risk tier classification, with policy_ref encoding the UBI policy identifier for state DOI UBI rate filing compliance audit and NAIC AVIWG guidance conformance documentation. Call with InsuranceAIContext.CLAIMS_DAMAGE for Lemonade AI or ClaimSearch AI claims damage assessment photograph ingestion before AI fraud indicator classification and payment authorisation determination, with inspection_id set to the claim number for 18 USC §1033 insurance fraud compliance documentation and NAIC Model 750 fraud prevention audit trail purposes. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four insurance underwriting and actuarial AI injection surfaces; covers how adversarial pixel-level perturbations cause AI misclassification without detectable visual artefacts at human review resolution, directly applicable to EagleView aerial inspection photograph and claims damage photograph injection.
• Vision-language model security — technical architecture of adversarial image attacks against vision-language models including pixel perturbation classes applicable to catastrophe model visualisation screenshot injection and claims damage assessment photograph manipulation in insurance AI workflows.
• Manufacturing quality inspection AI prompt injection — industrial AI injection context covering the broader class of physical inspection photograph manipulation applicable to property aerial inspection AI injection and claims damage assessment photograph adversarial attacks.
• Free tier — 10 scans/day, no card required — start scanning insurance underwriting and actuarial AI images at development volumes before committing to a production plan.