Prompt injection in retail and consumer banking AI

Four adversarial injection surfaces in retail and consumer banking AI

1. Mortgage loan document photograph injection (JPMorgan Chase COiN AI, Wells Fargo Fargo AI, Blend AI)

Mortgage loan document photograph AI processes scan images of borrower income statements, pay stub photographs, W-2 and tax return document scans, bank statement images, gift letter document photographs, and asset verification statement scans submitted through mortgage application portals, loan officer document upload interfaces, and digital mortgage origination platforms that extract income figures, employment verification data, asset balance totals, and debt-to-income ratio inputs from borrower-submitted document image inputs, generating income verification determinations, TRID disclosure discrepancy flags, debt-to-income ratio breach alerts, and ability-to-repay compliance classifications that govern whether mortgage loan officers proceed to loan commitment, require additional documentation from applicants, or flag the application for compliance review under TILA-RESPA Integrated Disclosure (TRID) and CFPB Regulation Z/X requirements. JPMorgan Chase COiN AI processes commercial and mortgage loan document scans through an AI contract intelligence and document extraction platform that eliminated 360,000 hours of annual legal review labor and is embedded in JPMorgan Chase’s mortgage origination workflow to extract income data from borrower document images, identify TRID disclosure discrepancies between loan estimate and closing disclosure figures, and generate income verification anomaly flags that mortgage compliance officers rely upon before loan commitment decisions. Wells Fargo Fargo AI processes mortgage origination document submissions through AI-assisted document review tools that support mortgage underwriting workflows, with Fargo AI’s document-reading capabilities embedded in Wells Fargo’s mortgage origination platform to surface income documentation anomalies, debt-to-income ratio breach conditions, and TRID compliance discrepancies that require mortgage officer review. Blend AI processes mortgage origination document submissions at more than 300 financial institution deployments, with Blend’s AI document ingestion and data extraction pipeline feeding underwriting decision support tools at Wells Fargo, US Bank, BMO, and regional mortgage originators operating under TRID regulatory disclosure timeline constraints that impose mandatory timing requirements on loan estimate and closing disclosure generation following document submission.

The adversarial injection surface is the borrower-submitted income statement scan, pay stub photograph, W-2 document image, bank statement scan, and asset verification document image submission pathway: mortgage loan supporting document images submitted through JPMorgan Chase COiN AI, Wells Fargo Fargo AI, or Blend AI document ingestion interfaces for AI income extraction, TRID discrepancy detection, and ability-to-repay classification. An adversarially crafted income statement scan — in which pixel perturbations applied to the gross income figure display region, employer identification header, or pay period date field on a borrower pay stub photograph cause the AI document extraction model to read an inflated income value when the actual document image records a lower income figure, or in which perturbations applied to the fee disclosure table region of a closing disclosure scan cause the AI TRID discrepancy detection tool to classify matching loan estimate and closing disclosure figures as within-tolerance when the actual document records a fee variance exceeding TRID tolerance thresholds — can suppress an income verification anomaly or TRID disclosure discrepancy flag that would otherwise mandate compliance officer review before loan commitment. In high-volume mortgage origination environments where Blend AI processes thousands of borrower document submissions per day across its financial institution customer base, a single adversarially manipulated closing disclosure scan that suppresses a TRID fee tolerance breach flag can allow a mortgage loan to proceed to closing without the compliance review that TILA-RESPA regulations require, creating regulatory liability for the originating institution under CFPB Regulation Z and Regulation X enforcement frameworks and potential False Claims Act exposure in Federal Housing Administration (FHA) or Veterans Administration (VA) guaranteed loan programs where income misrepresentation constitutes a material false statement to a federal agency.

The regulatory consequences of adversarially suppressed mortgage document AI detection span TILA-RESPA Integrated Disclosure enforcement, CFPB supervisory examination, Dodd-Frank ability-to-repay liability, and False Claims Act dimensions of exceptional severity for originating institutions. TRID (the TILA-RESPA Integrated Disclosure rule, 12 CFR Part 1026/Regulation Z and 12 CFR Part 1024/Regulation X) requires that mortgage originators provide borrowers with a Loan Estimate within three business days of application and a Closing Disclosure three business days before consummation, with specified fee tolerance thresholds governing permissible variance between Loan Estimate and Closing Disclosure line items; adversarial AI manipulation of closing disclosure document scan processing that suppresses a fee tolerance breach flag allows a TRID-violating loan to proceed to consummation without the required corrected disclosure and waiting period, creating a Regulation Z disclosure violation with CFPB civil money penalty exposure and potential loan rescission liability. Dodd-Frank Wall Street Reform Act §1411 (codified at 15 USC §1639c, Ability-to-Repay) requires mortgage originators to make a reasonable, good-faith determination of a borrower’s ability to repay before extending a covered mortgage loan, with the determination required to be based on verified and documented income and financial information; adversarial manipulation of AI income document extraction that causes an inflated income figure to be used in the ability-to-repay determination creates an ATR compliance failure with CFPB enforcement and private right of action litigation exposure. The False Claims Act (31 USC §§3729–3733) imposes treble damages and civil penalties on persons who knowingly submit false claims to the federal government; in FHA-insured or VA-guaranteed mortgage programs, adversarially manipulated income documentation that causes AI underwriting systems to recommend loan approval based on false income data may constitute a material false statement to HUD or the VA under the False Claims Act, with originating institution liability dimensions that substantially exceed standard CFPB civil money penalty exposure. Threshold: 60 for mortgage loan document photograph injection.

2. Bank statement and transaction document scan injection (Featurespace ARIC AI, NICE Actimize AI)

Bank statement and transaction document scan AI processes photograph submissions of multi-month bank statement documents, account transaction history scans, wire transfer record images, cash transaction receipt photographs, and business bank account statement scans submitted through AML monitoring document import channels, know-your-customer (KYC) document review portals, and bank statement analysis interfaces that extract transaction pattern classifications, cash deposit frequency and structuring indicators, overdraft behavioral patterns, and cash-intensive business anomaly scores from customer-submitted bank statement document image inputs, generating AML alert triage classifications, FinCEN Suspicious Activity Report (SAR) filing recommendation flags, OFAC transaction screening match indicators, and customer risk rating updates that govern whether compliance officers file SAR reports with FinCEN, escalate customer relationships for enhanced due diligence, or restrict account activity pending compliance review. Featurespace ARIC AI is deployed at HSBC, NatWest, Worldpay, Vocalink, and other major financial institutions as an adaptive behavioral analytics engine that processes transaction data streams and bank statement document submissions through Bayesian machine learning models specifically engineered to detect structuring transaction patterns (smurfing — the practice of dividing large cash transactions into smaller amounts designed to evade the Bank Secrecy Act’s §5313 Currency Transaction Report threshold of $10,000) and cash-intensive business anomalies that indicate potential money laundering. NICE Actimize AI is the leading AML transaction monitoring and financial crime compliance AI platform deployed at major global financial institutions including Barclays, BNY Mellon, Deutsche Bank, and hundreds of regional banks and credit unions, processing transaction monitoring alert feeds, bank statement documents, and SAR narrative drafting inputs through AI-assisted AML alert triage, suspicious activity detection, and regulatory filing decision support tools that directly govern FinCEN SAR filing obligation satisfaction under Bank Secrecy Act 31 USC §5318(g). Temenos AI’s core banking document processing capabilities are embedded in the banking operations of more than 3,000 financial institutions, where they process customer bank statement submissions and transaction record images through core banking workflow AI tools that feed AML monitoring and transaction surveillance systems with structured data extracted from document scan inputs.

The adversarial injection surface is the multi-month bank statement photograph, account transaction record scan, cash transaction receipt image, and business bank account statement document image submission pathway: bank statement and transaction record document images submitted through Featurespace ARIC AI or NICE Actimize AI AML document analysis interfaces for AI structuring pattern detection, cash-intensive business anomaly classification, and SAR filing obligation determination. An adversarially crafted bank statement photograph — in which pixel perturbations applied to the cash deposit amount display fields, transaction date column values, or account balance running-total figures on a multi-month bank statement document image cause the AML AI to classify a series of sub-$10,000 cash deposits — occurring on consecutive days, in amounts that collectively pattern into a structuring sequence — as non-anomalous retail deposit activity rather than a structuring (smurfing) indicator that would mandate a FinCEN SAR filing under 31 USC §5318(g), or in which perturbations applied to the business type and cash deposit frequency fields of a business bank account statement scan cause the AI to downgrade a cash-intensive business risk score that would otherwise trigger enhanced due diligence review — can suppress an AML structuring alert that would otherwise trigger a mandatory SAR filing obligation, allowing a money laundering pattern to continue without the FinCEN reporting that Bank Secrecy Act compliance requires. In AML document analysis environments where NICE Actimize AI processes millions of transaction records and bank statement documents annually across its global financial institution customer base, the adversarial suppression of structuring pattern detection in a single bank statement submission can allow a structuring campaign spanning multiple months and dozens of transactions to avoid FinCEN SAR filing while NICE Actimize AI’s AI alert triage system continues to classify the account’s transaction pattern as within the AI’s normal-activity behavioral model.

The regulatory consequences of adversarially suppressed AML document AI detection span Bank Secrecy Act enforcement, FinCEN SAR filing obligation, OFAC screening compliance, and criminal money laundering facilitation liability dimensions of exceptional severity for financial institutions. Bank Secrecy Act 31 USC §5318(g) imposes a mandatory obligation on financial institutions to file Suspicious Activity Reports with FinCEN when the institution knows, suspects, or has reason to suspect that a transaction involves funds from illegal activity, is designed to evade any reporting requirement, or lacks a lawful purpose; adversarial AI manipulation of bank statement transaction analysis that suppresses a structuring pattern classification eliminates the AI-generated SAR filing trigger that the institution’s compliance program relies upon, creating a SAR non-filing that exposes the institution to FinCEN civil money penalty liability of up to $1,000,000 per day per violation under 31 USC §5321 and potential criminal referral under 31 USC §5322. FinCEN’s SAR filing rules (31 CFR §1020.320 for banks) require SAR filings within 30 days of initial detection of a suspicious transaction or activity; adversarial suppression of an AI-detected structuring pattern in a bank statement document analysis workflow that causes a SAR filing to occur outside the 30-day detection window — or not at all — creates a SAR timeliness compliance failure with FinCEN examination consequences. FinCEN guidance on AML program requirements (FIN-2014-R007 and related SAR activity review documents) has consistently emphasized that financial institutions’ AML programs must include monitoring of document-based transaction review workflows; adversarial manipulation of AI document analysis that disables document-based structuring detection creates an AML program adequacy deficiency that bank regulatory examiners (OCC, Federal Reserve, FDIC) assess as a component of AML program safety-and-soundness examination. OFAC (Office of Foreign Assets Control) transaction screening obligations under the International Emergency Economic Powers Act (50 USC §§1701–1707) require financial institutions to block transactions involving OFAC-designated parties; adversarial bank statement document manipulation that suppresses AI-generated OFAC screening match indicators creates OFAC civil penalty exposure. Threshold: 60 for bank statement and transaction document scan injection.

3. Consumer loan application document injection (Zest AI, Upstart AI)

Consumer loan application document AI processes photograph submissions of borrower income documentation including pay stub scan images, W-2 form photographs, 1099 document scans, employer verification letter images, and bank statement document photographs submitted through consumer lending application portals, digital loan origination interfaces, and automated underwriting platform document upload channels that extract verified income figures, employment status classifications, income-to-debt-service ratio calculations, and thin-file risk indicator values from borrower-submitted income documentation image inputs, generating loan approval recommendations, credit risk tier classifications, adverse action flag triggers, and income verification anomaly alerts that govern whether lending platforms extend loan offers, generate adverse action notices under ECOA Regulation B, or escalate applications for human underwriter review under FCRA compliance requirements. Zest AI is deployed at more than 70 lender customers including community banks, credit unions, and regional financial institutions as a machine learning underwriting platform whose AI models process income documentation photographs and employment verification document images to generate credit decisioning recommendations that affect consumer loan approval rates, credit risk pricing, and adverse action determination workflows under the Equal Credit Opportunity Act (ECOA) and Fair Credit Reporting Act (FCRA) regulatory frameworks applicable to consumer credit decisions. Upstart AI is deployed at banking partners and credit unions including Customers Bank, First Federal Savings and Loan, and CBNA as an AI-powered income and employment verification platform that processes borrower-uploaded pay stub photographs, bank statement scan images, and employer letter document images through AI income verification models that assess whether applicant income documentation supports the income figures used in Upstart’s AI credit decision models, with Upstart’s income verification AI output directly feeding the credit decisioning recommendations that Upstart forwards to its banking and credit union lending partners for final loan commitment decisions. Ocrolus AI has processed more than one billion document pages for lending use cases and is deployed at fintech lenders, mortgage originators, and community banks to extract structured income, employment, and asset data from borrower-submitted document scan images including bank statements, pay stubs, tax returns, and profit-and-loss statements, with Ocrolus AI’s extracted data fields feeding underwriting decision systems at SoFi, BlueVine, Kabbage, and other digital lending platforms whose credit decisions are subject to ECOA and FCRA adverse action disclosure requirements.

The adversarial injection surface is the borrower-submitted pay stub photograph, W-2 form scan image, employer verification letter document image, and bank statement photograph submission pathway: income documentation images submitted through Zest AI or Upstart AI consumer lending underwriting platform document interfaces for AI income extraction, thin-file risk scoring, and adverse action flag trigger determination. An adversarially crafted pay stub photograph — in which pixel perturbations applied to the year-to-date gross earnings display field, hourly wage rate line, or employer tax identification number region on a borrower pay stub scan image cause the consumer lending AI to extract an income figure that places the applicant’s income-to-debt-service ratio below the lending platform’s adverse action threshold when the actual document image records a lower income that would require the platform to generate an ECOA Regulation B adverse action notice — can suppress an adverse action flag trigger that the ECOA and FCRA require the lending institution to act upon, allowing a loan approval recommendation to be generated based on AI-extracted income data that does not match the actual borrower income document, without generating the adverse action notice that ECOA Regulation B §202.9 mandates when credit is denied or offered on terms materially different from those requested. In automated underwriting platform environments where Zest AI processes underwriting recommendations across its 70+ lender customer base without human income verification officer review of each individual document image, adversarial manipulation of a pay stub photograph that inflates the AI-extracted income figure beyond the platform’s adverse action threshold eliminates the adverse action notice obligation while also embedding a false income basis in the credit risk model’s decisioning record, creating downstream FCRA compliance complications when the loan subsequently defaults and the lending institution conducts credit decision audit review.

The regulatory consequences of adversarially suppressed consumer loan application document AI detection span ECOA Regulation B adverse action notice obligations, FCRA adverse action disclosure requirements, CFPB supervisory examination authority, and consumer protection enforcement dimensions applicable to consumer lending institutions operating under federal fair lending statutes. ECOA (Equal Credit Opportunity Act, 15 USC §1691 et seq.) and its implementing Regulation B (12 CFR Part 202) require creditors to provide adverse action notices to credit applicants when credit is denied, terminated, or offered on terms less favorable than those requested or applied for, with the notice required to state the specific reasons for the adverse action or disclose the applicant’s right to obtain those reasons; adversarial AI manipulation of income documentation processing that suppresses an adverse action flag trigger — allowing a loan denial or counter-offer to occur without an ECOA-compliant adverse action notice — creates a Regulation B §202.9 violation with CFPB civil money penalty exposure and private right of action litigation liability of up to $10,000 per violation under 15 USC §1691e. The Fair Credit Reporting Act (15 USC §1681m) imposes adverse action disclosure requirements on creditors who take adverse action based wholly or partly on information contained in a consumer report, including the requirement to identify the consumer reporting agency and provide specific reasons for the adverse action; adversarial manipulation of AI income documentation processing that affects the credit decision basis creates FCRA adverse action disclosure compliance complications with Federal Trade Commission and CFPB enforcement authority. CFPB Regulation B (12 CFR §202.9) requires adverse action notices to be provided within 30 days of receiving a completed application; adversarial income document manipulation that causes an AI underwriting platform to recommend loan approval on the basis of inflated income data — which is later discovered through post-origination income verification — can create adverse action notice timeliness failures when the lending institution then denies or modifies the loan after the 30-day ECOA window has elapsed. The CFPB’s fair lending examination procedures (CFPB Examination Procedures § Fair Lending) evaluate whether creditors’ automated underwriting systems apply consistent credit standards across protected class and non-protected class applicants; adversarial income document manipulation that systematically affects credit decisions within an AI underwriting platform creates disparate impact analysis complications in CFPB fair lending examination. Threshold: 60 for consumer loan application document injection.

4. Check and payment instrument image injection (bank AI fraud detection systems)

Check and payment instrument image AI processes photographs and scan images of paper checks, money orders, cashier’s check documents, electronic check (e-check) images captured through remote deposit capture (RDC) channels, and substitute check images processed through bank clearing systems submitted through mobile deposit capture interfaces, branch scanner deposit channels, and commercial lockbox check processing portals that extract payee name fields, MICR line routing and account number values, check amount figures, check date fields, and instrument authentication features from check image inputs, generating altered payee detection flags, duplicate presentment indicators, MICR line anomaly classifications, and counterfeit instrument pattern scores that govern whether bank fraud detection AI systems clear check payments for posting to recipient accounts, place exception holds pending manual fraud review, or return instruments as unpayable. Featurespace ARIC AI’s fraud detection capabilities include check fraud pattern detection applied to transaction-level check processing behavioral data and check image metadata, with ARIC AI’s adaptive behavioral models identifying duplicate presentment patterns, altered payee behavioral signatures, and account-level check fraud anomaly indicators across the transaction streams of HSBC, NatWest, and other major financial institution deployments. NICE Actimize AI’s check fraud detection platform processes check image submissions and remote deposit capture images through AI-assisted fraud scoring models that generate altered payee detection flags, duplicate presentment alerts, and counterfeit check pattern classifications for financial institution fraud operations teams at Barclays, BNY Mellon, Deutsche Bank, and global banking institution deployments. Temenos AI’s core banking payment processing AI handles check payment processing workflows at more than 3,000 financial institution deployments globally, with Temenos AI payment processing tools generating fraud risk classifications and exception holds based on AI analysis of check image submissions and payment instrument data extracted from deposit capture scan inputs. Bank of America Erica AI’s account monitoring capabilities include AI-assisted transaction surveillance that processes check deposit activity and remote deposit capture events through behavioral anomaly detection models that flag suspicious deposit patterns for fraud review by Bank of America fraud operations teams.

The adversarial injection surface is the mobile remote deposit capture check photograph, branch scanner check image, commercial lockbox check scan, and substitute check image submission pathway: check and payment instrument scan images submitted through bank AI fraud detection interfaces for AI altered payee detection, duplicate presentment identification, MICR line anomaly classification, and counterfeit instrument pattern scoring. An adversarially crafted check photograph — in which pixel perturbations applied to the payee name line region, check amount written-amount field, or instrument security feature watermark area on a remote deposit capture check image cause the AI fraud detection system to classify an altered payee — where the original payee name has been chemically or digitally altered to a fraudulent payee name — as an unaltered authentic instrument, suppressing the altered payee flag that would otherwise cause the bank’s fraud operations team to place a hold on the check and initiate presentment review, or in which perturbations applied to the MICR encoding line and check number field on a check scan image cause a duplicate presentment indicator to be suppressed when the same check image was previously presented for deposit at a different institution — can suppress fraud detection alerts that bank fraud operations depend upon to satisfy UCC Article 4 check presentment warranty obligations and Regulation CC availability-of-funds compliance requirements. In mobile banking environments where remote deposit capture processes millions of check images daily through AI fraud scoring pipelines that operate at the sub-second latency required for real-time availability determination, adversarial suppression of an altered payee flag or duplicate presentment indicator in a check image fraud scoring pipeline allows fraudulent instruments to be cleared for fund availability before fraud operations teams can intervene, with the associated funds potentially withdrawn before the bank’s next-day fraud review cycle identifies the fraudulent presentment.

The regulatory and legal consequences of adversarially suppressed check and payment instrument AI fraud detection span UCC Article 4 check presentment warranty obligations, Regulation CC availability of funds compliance requirements, OCC fraud guidance, and Federal Reserve check processing rules with direct financial institution loss and regulatory liability dimensions. Uniform Commercial Code Article 4 (Bank Deposits and Collections) imposes presentment warranties on banks that transfer check instruments through the collection process, including the warranty that the instrument has not been altered; a bank whose AI fraud detection system fails to identify an altered check — due to adversarial image manipulation — and clears the altered check for payment may face UCC Article 4-208 warranty breach liability when the paying bank returns the altered instrument, with the collecting bank bearing the cost of the warranty breach if it cannot demonstrate that it exercised good faith and ordinary care in the check clearing process. Regulation CC (Availability of Funds and Collection of Checks, 12 CFR Part 229) governs funds availability schedules for deposited checks and imposes obligations on depositary banks to make funds from deposited checks available within specified timeframes; adversarial suppression of a duplicate presentment or counterfeit check indicator that causes a fraudulent check to be cleared and funds made available under Regulation CC’s next-day or second-day availability schedule creates a funds availability compliance scenario where the bank has satisfied Regulation CC’s customer-facing availability obligation while simultaneously failing to detect fraud — with the resulting consumer funds withdrawal creating a net loss position for the depositary bank that cannot be recovered through Regulation CC chargeback procedures if the funds have been withdrawn and the underlying check is subsequently dishonored. OCC fraud guidance (OCC Bulletin 2019-37, Bank Fraud — Sound Practices) directs nationally chartered banks to implement layered fraud controls that include image-based check fraud detection; adversarial manipulation of AI check image fraud detection that systematically suppresses altered payee or duplicate presentment flags creates an OCC fraud control adequacy deficiency with bank examination consequences. The Federal Reserve’s Check 21 rules (12 CFR Part 229, Subpart D) govern substitute check creation and processing; adversarial check image manipulation targeting the substitute check image pathway creates Federal Reserve operational check processing rule compliance concerns with Board of Governors enforcement authority. Threshold: 60 for check and payment instrument image injection.

Integration: retail and consumer banking AI document image ingestion with Glyphward pre-scan

Retail and consumer banking AI document image ingestion flows from mortgage application document upload portals, AML bank statement import channels, consumer loan application document submission interfaces, and remote deposit capture check image pipelines into mortgage document AI, AML transaction surveillance AI, consumer lending underwriting AI, and payment instrument fraud AI processing layers. Insert Glyphward’s pre-scan at the ingestion boundary before AI-generated output is committed to mortgage underwriting records, AML alert queues, adverse action determination logs, or check clearing decisions:

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"

# Retail & consumer banking AI — TILA-RESPA TRID, CFPB Regulation Z/X,
# BSA 31 USC §5318, ECOA Regulation B, FCRA 15 USC §1681, UCC Article 4,
# Regulation CC, OCC fraud guidance, Dodd-Frank §1411 ability-to-repay.
# Suppression of TRID discrepancy flags, AML structuring indicators,
# adverse action triggers, and counterfeit instrument patterns creates
# regulatory enforcement liability across CFPB, FinCEN, OCC, and Federal Reserve.
THRESHOLD_BANKING = 60  # uniform across all four banking AI contexts


class BankingAIContext(str, Enum):
    MORTGAGE_DOCUMENT   = "mortgage_document"    # JPMorgan COiN, Wells Fargo Fargo, Blend
    BANK_STATEMENT      = "bank_statement"        # Featurespace ARIC, NICE Actimize
    LOAN_APPLICATION    = "loan_application"      # Zest AI, Upstart AI
    PAYMENT_INSTRUMENT  = "payment_instrument"    # bank check fraud AI (ARIC, Actimize, Temenos)


def threshold_for(context: BankingAIContext) -> int:
    # All four retail and consumer banking AI surfaces carry the same
    # threshold of 60: TRID disclosure suppression, BSA SAR non-filing,
    # ECOA adverse action failure, and UCC Article 4 warranty breach each
    # carry comparable regulatory enforcement severity.
    return THRESHOLD_BANKING


async def scan_banking_ai_image(
    image_path: str | Path,
    context: BankingAIContext,
    institution_id_hash: str,  # SHA-256 of financial institution routing number or LEI
    loan_ref: str,             # e.g. "MTG-2026-00771", "CONS-2026-44821", "CHK-DEP-9923"
    document_scan_id: str,     # e.g. document upload UUID from mortgage platform or RDC system
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan a retail or consumer banking AI document image for adversarial
    injection payloads before forwarding to mortgage document AI, AML bank
    statement AI, consumer lending underwriting AI, or payment instrument
    fraud AI processing systems.

    Raises AdversarialBankingAIImageError if score meets threshold:
      - MORTGAGE_DOCUMENT:   threshold 60; TILA-RESPA TRID; CFPB Regulation Z/X;
                             Dodd-Frank §1411 ATR; False Claims Act (FHA/VA)
      - BANK_STATEMENT:      threshold 60; BSA 31 USC §5318 AML program;
                             FinCEN SAR 31 CFR §1020.320; OFAC screening
      - LOAN_APPLICATION:    threshold 60; ECOA Regulation B §202.9 adverse action;
                             FCRA 15 USC §1681m; CFPB fair lending examination
      - PAYMENT_INSTRUMENT:  threshold 60; UCC Article 4-208 presentment warranty;
                             Regulation CC 12 CFR Part 229; OCC Bulletin 2019-37;
                             Federal Reserve Check 21 rules
    """
    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": {
                "banking_context":    context.value,
                "institution_id_hash": institution_id_hash,
                "loan_ref":           loan_ref,
                "document_scan_id":   document_scan_id,
                "client_scan_id":     client_scan_id,
                "image_sha256":       image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "institution_id_hash": institution_id_hash,
        "loan_ref":            loan_ref,
        "document_scan_id":    document_scan_id,
        "banking_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_banking_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialBankingAIImageError(
            f"Banking AI document image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"institution={institution_id_hash} loan_ref={loan_ref}"
        )
    return result


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


class AdversarialBankingAIImageError(Exception):
    """Raised when a retail or consumer banking AI document image exceeds the adversarial injection threshold."""
    pass

Call scan_banking_ai_image() with BankingAIContext.MORTGAGE_DOCUMENT before forwarding borrower income statement scans, pay stub photographs, and closing disclosure document images to JPMorgan Chase COiN AI, Wells Fargo Fargo AI, or Blend AI document ingestion pipelines — the highest TRID compliance integration point, where adversarial suppression of a fee tolerance breach flag or income verification anomaly defers required compliance officer review past TRID disclosure deadlines and exposes the originating institution to Regulation Z civil money penalty liability. Call with BankingAIContext.BANK_STATEMENT for bank statement photographs and transaction record scan images before forwarding to Featurespace ARIC AI or NICE Actimize AI AML document analysis interfaces, preserving image_sha256 as the forensic anchor for FinCEN SAR filing obligation audit trail documentation that demonstrates a technical control was in place at the document submission boundary if structuring pattern suppression is alleged in an OCC or FinCEN AML examination. Call with BankingAIContext.LOAN_APPLICATION for Zest AI or Upstart AI consumer lending income documentation image submissions before AI income extraction and adverse action flag determination, with loan_ref encoding the application reference number for ECOA Regulation B adverse action notice audit trail and CFPB fair lending examination documentation. Call with BankingAIContext.PAYMENT_INSTRUMENT for remote deposit capture check photographs, branch scanner check images, and commercial lockbox check scans before AI altered payee detection, duplicate presentment scoring, and counterfeit instrument classification, with document_scan_id linking the Glyphward scan record to the specific remote deposit capture session identifier for UCC Article 4 ordinary care and Regulation CC funds availability compliance audit trail purposes. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four retail and consumer banking AI injection surfaces; covers how adversarial pixel-level perturbations cause AI document extraction misclassification without producing detectable visual artifacts at the resolution of human compliance officer document review.
• Vision-language model security — technical architecture of adversarial image attacks against vision-language and document-reading AI models including gradient-based pixel perturbation techniques applicable to mortgage document AI income extraction injection and AML bank statement analysis suppression.
• Multimodal AI security testing — covers the broader class of document image manipulation attacks applicable to consumer banking AI document ingestion pipelines, including test methodologies for evaluating adversarial robustness in income documentation AI and check fraud detection AI.
• Prompt injection for financial services AI — financial services AI security context covering adversarial injection across the full spectrum of banking, lending, and payments AI platforms with CFPB, OCC, FinCEN, and Federal Reserve regulatory consequence analysis.
• Free tier — 10 scans/day, no card required — start scanning retail and consumer banking AI document images at development volumes before committing to a production integration plan aligned to TRID, BSA, or ECOA compliance audit trail requirements.