Prompt injection in government and public sector AI

Four adversarial injection surfaces in government and public sector AI

1. Benefits eligibility document injection (Tyler Technologies AI, Salesforce Government Cloud AI)

Benefits eligibility document AI processes citizen application document photographs and eligibility verification scan images from Tyler Technologies NIC AI across more than 15,000 government agency customers including county SNAP and Medicaid offices, housing assistance bureaus, and state human services departments; Salesforce Government Cloud AI at state human services agencies managing SNAP, Medicaid, TANF, and Section 8 housing assistance caseloads; IBM Watsonx Government at federal benefit programme operations including HHS, USDA FNS, and HUD grant and benefits administration; and Granicus GovDelivery AI at state benefit notification and case management operations processing beneficiary communication and eligibility status record images — extracting benefits eligibility and qualification indicator classifications from citizen application document photograph inputs in AI-assisted eligibility determination and case management pipelines, generating eligibility approval or denial recommendations, caseworker review priority assignments, benefit amount calculation inputs, and eligibility documentation records that government caseworkers and benefits administrators depend upon for SNAP 7 USC §2014 eligibility determination compliance, Medicaid 42 USC §1396a state plan requirements, Section 8 Housing Choice Voucher programme 42 USC §1437f eligibility obligations, and ADA §504 non-discriminatory programme access requirements at case volumes that make individual human caseworker review of every AI-processed application document impracticable. Tyler Technologies NIC AI’s benefits eligibility platform processes citizen application photographs through AI-assisted eligibility indicator extraction and programme qualification classification tools that county benefits offices use for SNAP, Medicaid, TANF, and housing assistance programme eligibility determination at caseloads of thousands of monthly applications where AI-assisted processing is operationally necessary for timely eligibility determination within statutory time limits. Salesforce Government Cloud AI’s case management tools process beneficiary eligibility document images through AI-assisted case eligibility and programme qualification classification tools at state human services agencies where AI-generated eligibility determinations feed caseworker review workflows and benefit award calculations.

The adversarial injection surface is the citizen benefits eligibility document photograph submission pathway: Tyler Technologies AI or Salesforce Government Cloud AI benefits eligibility document scan images submitted through AI-assisted eligibility indicator classification and programme qualification assessment tools for AI eligibility determination recommendation and caseworker review priority assignment. An adversarially crafted SNAP or Medicaid application document photograph — in which pixel perturbations applied to the income verification document display region, the household composition indicator visual marker, or the resource and asset documentation display in a benefits application document photograph cause the AI to classify an applicant satisfying programme eligibility criteria under 7 USC §2014 SNAP income and resource standards or 42 USC §1396a Medicaid income eligibility thresholds as a below-threshold ineligible applicant when the actual document photograph evidences eligibility meeting programme qualification criteria — can suppress an eligibility indicator that would otherwise generate a benefits approval recommendation, a caseworker positive determination input, and an award calculation initiation record. In county benefits office environments where Tyler Technologies AI or Salesforce Government Cloud AI processes thousands of monthly application documents without individual caseworker pixel-level examination of every AI-processed document before the AI eligibility classification governs the caseworker’s determination workflow, adversarial suppression of eligibility indicators creates systematic benefits denial consequences for eligible citizens that trigger 42 USC §1983 civil rights due process deprivation liability and ADA §504 non-discriminatory programme access obligations for government agencies administering federally assisted benefit programmes.

The 42 USC §1983 civil rights and due process consequences of adversarially suppressed eligibility classification in benefits eligibility AI span constitutional due process, ADA §504, Goldberg v. Kelly pre-deprivation hearing, and Mathews v. Eldridge due process balancing dimensions. 42 USC §1983 provides a civil cause of action against persons acting under colour of state law who deprive citizens of rights, privileges, or immunities secured by the Constitution or federal statutes; a government agency’s use of an adversarially corrupted Tyler Technologies AI or Salesforce Government Cloud AI benefits eligibility determination that denies SNAP, Medicaid, or housing assistance to an eligible citizen creates a colour-of-state-law deprivation of the citizen’s programme eligibility rights with §1983 liability dimensions for the agency. Goldberg v. Kelly, 397 US 254 (1970), established that welfare benefits recipients are entitled to an evidentiary hearing before benefits are terminated; the Supreme Court held that welfare benefits constitute statutory entitlements protected by due process — adversarial manipulation of benefits eligibility AI that generates erroneous eligibility denial recommendations without the pre-deprivation procedural protections required by Goldberg creates constitutional due process violation exposure. Mathews v. Eldridge, 424 US 319 (1976), established a three-factor balancing test for determining what process is due before government deprivation of a protected interest — weighing the private interest affected, the risk of erroneous deprivation with existing procedures, and the government’s interest — adversarial manipulation of benefits eligibility AI that systematically increases the risk of erroneous deprivation through corrupted AI classification shifts the Mathews balance toward requiring more robust procedural protections than AI-only eligibility determination provides. ADA §504 of the Rehabilitation Act prohibits discrimination based on disability in federally assisted programmes; adversarial manipulation of benefits eligibility AI that disproportionately suppresses eligibility indicators in applications submitted by disabled citizens creates ADA §504 discriminatory programme access dimensions. Threshold: 55 for benefits eligibility AI — reflecting the 42 USC §1983 civil rights, ADA §504, Goldberg v. Kelly pre-deprivation, and Mathews v. Eldridge due process dimensions of adversarially manipulated eligibility classification.

2. Law enforcement intelligence AI injection (Palantir Gotham AI, Motorola Solutions CommandCentral AI)

Law enforcement intelligence AI processes criminal network link analysis visualisation displays, pattern-of-life intelligence synthesis images, criminal record and investigation history displays, and intelligence fusion analysis visualisation outputs from Palantir Gotham AI at US and UK intelligence agencies, CBP, ICE, HSI, NYPD, and LAPD law enforcement operations; Motorola Solutions CommandCentral AI at more than 1,000 law enforcement agency deployments across the United States; Palantir AIP Government at federal law enforcement and intelligence agency operations integrating AI-assisted analysis of multimodal intelligence data inputs; and i2 Analyst’s Notebook AI at financial intelligence, law enforcement, and national security agency link analysis programme deployments — extracting criminal network indicator classifications and investigative priority determinations from intelligence link analysis visualisation display inputs in AI-assisted law enforcement decision support and investigative targeting pipelines, generating investigative focus recommendations, surveillance authorisation support inputs, enforcement action priority assignments, and criminal network connection probability assessments that intelligence analysts and law enforcement decision-makers depend upon for Fourth Amendment warrant requirement compliance, Brady v. Maryland prosecutorial evidence disclosure obligation management, and 18 USC §1519 records integrity compliance dimensions. Palantir Gotham AI’s intelligence analysis platform processes law enforcement and intelligence agency data through AI-assisted link analysis, network visualisation, and pattern recognition tools that intelligence analysts use to identify criminal network connections, investigative targets, and enforcement priority assessments from multimodal intelligence inputs including document images, surveillance images, and network visualisation displays at intelligence processing volumes that make individual human analyst review of every AI-processed intelligence image impracticable.

The adversarial injection surface is the law enforcement intelligence link analysis visualisation display submission pathway: Palantir Gotham AI or Motorola Solutions CommandCentral AI intelligence link analysis visualisation displays submitted through AI-assisted criminal indicator classification and investigative priority assessment tools for AI intelligence analysis recommendation and enforcement decision support. An adversarially crafted Palantir Gotham link analysis display — in which pixel perturbations applied to the criminal network connection visual indicator region, the subject node linkage probability display marker, or the pattern-of-life activity correlation visualisation in a Palantir Gotham intelligence network display image cause the AI to classify a criminal network display evidencing significant criminal indicator connections requiring investigative prioritisation as a below-threshold intelligence profile not meeting the investigative priority flag threshold when the actual visualisation documents criminal network connections meeting Palantir Gotham AI’s investigative priority classification criteria — can suppress an investigative priority flag that would otherwise generate a law enforcement investigative action recommendation, a surveillance authorisation support input, and a criminal network intelligence assessment record. Adversarial law enforcement intelligence AI injection that suppresses criminal indicator detections creates Brady v. Maryland prosecutorial disclosure obligation dimensions — where exculpatory or material evidence that the prosecution possesses but fails to disclose to the defence results in Brady violations — when AI-generated intelligence analysis that failed to surface criminal network connections adverse to the prosecution’s theory due to adversarial manipulation creates undisclosed material exculpatory evidence obligations.

The Fourth Amendment and Brady v. Maryland consequences of adversarially manipulated criminal indicator classification in law enforcement intelligence AI span Fourth Amendment exclusionary rule, Brady v. Maryland 373 US 83 prosecutorial disclosure, Giglio v. United States 405 US 150 witness credibility evidence, and 18 USC §1519 records obstruction dimensions. The Fourth Amendment exclusionary rule — established in Mapp v. Ohio, 367 US 643 (1961), as applicable to state law enforcement — requires suppression of evidence obtained through unconstitutional searches; adversarially manipulated Palantir Gotham AI intelligence analysis that generates inaccurate criminal network probability scores used to support warrant applications creates Fourth Amendment fruit-of-the-poisonous-tree suppression dimensions when adversarially corrupted AI outputs contributed to probable cause determinations. Brady v. Maryland, 373 US 83 (1963), established that the prosecution’s suppression of evidence material to guilt or punishment violates due process; adversarial manipulation of Palantir Gotham AI intelligence analysis that conceals exculpatory criminal network connection data — data that would have been material to the defence — creates Brady disclosure failure dimensions when prosecutors did not disclose the AI analysis output because adversarial manipulation caused the AI to suppress the material indicator. Giglio v. United States, 405 US 150 (1972), extended Brady obligations to evidence affecting witness credibility; adversarially manipulated intelligence AI that suppresses indicators affecting the reliability of government witness testimony creates Giglio disclosure failure dimensions. 18 USC §1519 criminalises knowingly falsifying, concealing, or covering up records or documents in connection with federal investigations; adversarial manipulation of law enforcement intelligence AI records creating false or misleading intelligence summaries for federal investigations creates §1519 obstruction of justice dimensions. Threshold: 60 for law enforcement intelligence AI — reflecting the Fourth Amendment exclusionary rule, Brady v. Maryland prosecutorial disclosure, Giglio witness credibility, and 18 USC §1519 obstruction dimensions of adversarially manipulated criminal indicator classification.

3. Permitting and licensing document injection (Tyler Technologies AI, NIC AI)

Permitting and licensing document AI processes building permit application document photographs, professional licence application document scans, driver licence and vehicle registration document images, business licence application document scans, and alcohol beverage control (ABC) licensing document photographs from Tyler Technologies AI permitting platform at more than 15,000 government agency customers including building and development services departments, state DMV operations, professional licensing boards including medical, legal, and engineering licensing authorities, and state ABC boards; NIC AI at state driver licensing, business licensing, and ABC licensing operations; and ESRI ArcGIS Pro AI at government geospatial permitting and land use management operations processing site plan and zoning compliance document images — extracting compliance indicator classifications and permitting eligibility determinations from permit and licence application document scan image inputs in AI-assisted permitting adjudication and licensing decision support pipelines, generating permit approval or denial recommendations, code compliance status assessments, licence eligibility determinations, and DMV eligibility qualification records that building officials, licensing board staff, and DMV examiners depend upon for building code compliance, professional licencee qualification, driver qualification, and business operating licence eligibility determination at application volumes that make individual human examiner review of every AI-processed permitting document impracticable. Tyler Technologies AI’s permitting platform processes building permit application documents through AI-assisted code compliance indicator extraction and permit eligibility classification tools that municipal building departments use for building permit application review at application volumes where AI-assisted processing is operationally necessary for meeting statutory permit review timelines.

The adversarial injection surface is the building permit, professional licence, DMV, and business licence application document scan image submission pathway: Tyler Technologies AI or NIC AI permitting and licensing document scan images submitted through AI-assisted compliance indicator classification and licence eligibility determination tools for AI permitting adjudication recommendation and licensing decision support. An adversarially crafted building permit application document photograph — in which pixel perturbations applied to the structural engineering compliance certification display region, the fire safety code compliance indicator visual marker, or the zoning setback compliance documentation display in a building permit application document scan cause the AI to classify a permit application meeting municipal building code compliance requirements and zoning ordinance specifications as a below-threshold non-compliant application triggering permit denial recommendation when the actual document scan evidences building code compliance meeting municipal permitting approval criteria — can suppress a compliance indicator that would otherwise generate a permit approval recommendation, a building official positive determination workflow input, and a permit issuance processing record. Building permit denials generated by adversarially corrupted Tyler Technologies AI permitting platform that suppress compliance indicators for compliant applications create 42 USC §1983 civil rights dimensions when the permit denial deprives the applicant of a constitutionally protected property interest in the permit without due process, and create state Administrative Procedure Act procedural compliance dimensions for agency action that must be supported by the administrative record. Professional licence application adversarial injection that suppresses qualification indicators for qualified applicants creates state professional licencee due process dimensions and 42 USC §1983 occupational licence liberty interest deprivation exposure.

The state Administrative Procedure Act, OSHA, 42 USC §1983, and professional licencee due process consequences of adversarially suppressed compliance classification in permitting and licensing AI span state APA administrative record requirements, municipal building code compliance, OSHA 29 CFR Part 1926 construction safety, professional licensure constitutional due process, and 42 USC §1983 property and liberty interest deprivation dimensions. State Administrative Procedure Acts — including the Model State APA, California APA Government Code §11340, New York State APA Executive Law §101, and Texas APA Government Code §2001 — require that agency adjudicative decisions be supported by the substantial evidence in the administrative record; adversarially corrupted Tyler Technologies AI permitting platform decisions that generate denial recommendations unsupported by the actual document evidence in the permit application record create state APA substantial evidence reviewability dimensions in administrative appeal proceedings. OSHA 29 CFR Part 1926 establishes construction industry safety standards including structural requirements, fire protection, and site safety; adversarially manipulated building permit AI that suppresses building code compliance indicators and delays compliant permit approvals creates downstream OSHA construction safety timeline dimensions when projects are delayed by permit denials predicated on adversarially corrupted AI compliance classification. NIC AI driver licence application document processing that suppresses applicant qualification indicators creates DMV due process dimensions and state administrative hearing rights for licence denial appeals under state motor vehicle code administrative hearing procedures. Threshold: 60 for permitting and licensing AI — reflecting the state APA, municipal building code, OSHA construction safety, professional licensure due process, and 42 USC §1983 property and liberty interest deprivation dimensions of adversarially manipulated compliance classification.

4. Election and voter registration document injection (voter registration AI)

Election and voter registration document AI processes voter registration application document scans, state ID and driver licence document photographs used for NVRA Motor Voter automatic registration, proof of citizenship documentation images, address verification document scans, and voter eligibility determination record images from Election Systems & Software (ES&S) AI at election administration operations across more than 40 US states; Hart InterCivic AI at county election administration operations; Dominion Voting Systems AI at election administration operations across more than 30 US states; Tyler Technologies elections division AI at county election administration operations; and state Secretary of State voter registration AI platforms processing DMV voter registration NVRA automatic registration document scan images — extracting voter eligibility indicator classifications and registration status determination inputs from voter registration document scan image inputs in AI-assisted voter registration processing and eligibility determination pipelines, generating voter registration approval or denial recommendations, eligibility status classification records, rolls maintenance action inputs, and voter registration documentation entries that election officials and county registrars depend upon for 52 USC §20501 NVRA voter registration rights obligation compliance, 52 USC §10101 Voting Rights Act compliance, and Help America Vote Act 52 USC §21082 provisional ballot and voter eligibility determination requirements at registration processing volumes that make individual human election official review of every AI-processed voter registration document impracticable. ES&S, Hart InterCivic, Dominion, and Tyler Technologies elections AI platforms collectively process voter registration and election administration data across the vast majority of US election jurisdictions, and their AI-assisted voter registration and eligibility determination tools operate at the critical interface between citizen voter registration applications and official voter rolls that determine electoral participation rights under the National Voter Registration Act and Voting Rights Act.

The adversarial injection surface is the voter registration application document scan and state ID or driver licence photograph submission pathway: ES&S AI, Hart InterCivic AI, Dominion Voting Systems AI, or state SOS voter registration AI platform document scan images submitted through AI-assisted voter eligibility indicator classification and registration status determination tools for AI voter registration processing and eligibility status classification. An adversarially crafted voter registration application document scan — in which pixel perturbations applied to the proof of citizenship documentation display region, the state residence address verification indicator visual marker, or the age eligibility documentation display in a voter registration application document scan cause the AI to classify a citizen satisfying voter eligibility requirements under state election code citizenship, residency, and age qualification criteria as a below-threshold ineligible applicant triggering registration denial when the actual document scan evidences voter eligibility meeting state election code qualification criteria — can suppress a voter eligibility indicator that would otherwise generate a voter registration approval recommendation, an election official positive registration processing input, and a voter rolls inclusion record. Adversarial voter registration document AI injection that systematically suppresses eligibility indicators for registration applicants from specific geographic areas, demographic groups, or registration submission channels creates 52 USC §20501 NVRA voter registration rights violation dimensions and 18 USC §594 voter intimidation and civil rights criminal exposure when adversarial manipulation of voter registration AI operates to functionally disenfranchise eligible voters through AI-generated registration denials.

The NVRA, Voting Rights Act, HAVA, 42 USC §1983, and 18 USC §594 consequences of adversarially manipulated voter eligibility classification in voter registration AI span 52 USC §20501 NVRA voter registration rights, 52 USC §10101 Voting Rights Act non-discrimination, HAVA 52 USC §21082 provisional ballot requirements, 42 USC §1983 constitutional voting rights, and 18 USC §594 voter intimidation dimensions. The National Voter Registration Act, 52 USC §20501 et seq., establishes the right of eligible citizens to register to vote through motor vehicle licensing agencies (Motor Voter), mail registration, and public assistance agencies; NVRA imposes affirmative obligations on states to facilitate voter registration and prohibits states from removing registered voters from rolls without meeting NVRA procedural requirements. Adversarially manipulated voter registration AI that systematically denies registration to eligible citizens through suppressed eligibility indicator classification creates NVRA enforcement dimensions — NVRA authorises the US Attorney General and state voter registration applicants to bring civil enforcement actions, and 52 USC §20511 imposes criminal penalties for knowing violations of NVRA provisions. The Voting Rights Act, 52 USC §10101, prohibits denial or abridgement of the right to vote based on race, colour, or previous condition of servitude and prohibits application of voting qualifications that result in denial of the right to vote; adversarial manipulation of voter registration AI that disproportionately suppresses registration approvals for minority voter registration applicants creates Voting Rights Act Section 2 enforcement dimensions. HAVA 52 USC §21082 requires states to provide provisional ballots to individuals who assert eligibility to vote but whose eligibility cannot be confirmed at the polls; adversarially corrupted voter registration AI that generates inaccurate voter rolls records creates HAVA provisional ballot obligation dimensions and post-election audit complications. Threshold: 65 for voter registration AI — reflecting the 52 USC §20501 NVRA voter registration rights, Voting Rights Act, HAVA provisional ballot, 42 USC §1983 constitutional voting rights, and 18 USC §594 voter intimidation dimensions of adversarially manipulated voter eligibility classification.

Integration: government and public sector AI image ingestion with Glyphward pre-scan

Government and public sector AI image ingestion flows from Tyler Technologies and Salesforce Government Cloud benefits eligibility document photograph channels, Palantir Gotham and Motorola Solutions CommandCentral law enforcement intelligence link analysis visualisation interfaces, Tyler Technologies and NIC permitting and licensing document scan platforms, and ES&S, Hart InterCivic, Dominion, and state SOS voter registration document scan AI processing systems into benefits eligibility determination AI, law enforcement investigative priority classification AI, permitting adjudication AI, and voter registration eligibility classification AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary before AI-generated output is committed to benefits eligibility recommendations, law enforcement investigative priority flags, permitting adjudication decisions, or voter registration status classifications:

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"

# Government & public sector AI — 42 USC §1983 civil rights due process;
# ADA §504 non-discrimination; Goldberg v. Kelly 397 US 254; Mathews v. Eldridge
# 424 US 319; Fourth Amendment exclusionary rule; Brady v. Maryland 373 US 83;
# 18 USC §1519 obstruction; 52 USC §20501 NVRA; HAVA 52 USC §21082.
THRESHOLD_BENEFITS_ELIGIBILITY_AI   = 55  # Tyler/Salesforce; §1983; ADA §504; Goldberg
THRESHOLD_LAW_ENFORCEMENT_INTEL_AI  = 60  # Palantir Gotham; 4th Amend; Brady; §1519
THRESHOLD_PERMITTING_LICENSING_AI   = 60  # Tyler/NIC; state APA; OSHA; §1983
THRESHOLD_VOTER_REGISTRATION_AI     = 65  # ES&S/Dominion; NVRA; VRA; HAVA; §594


class GovPublicSectorAIContext(str, Enum):
    BENEFITS_ELIGIBILITY_AI   = "benefits_eligibility_ai"   # Tyler, Salesforce Gov Cloud
    LAW_ENFORCEMENT_INTEL_AI  = "law_enforcement_intel_ai"  # Palantir Gotham, Motorola
    PERMITTING_LICENSING_AI   = "permitting_licensing_ai"   # Tyler, NIC, ESRI
    VOTER_REGISTRATION_AI     = "voter_registration_ai"     # ES&S, Dominion, Hart


def threshold_for(context: GovPublicSectorAIContext) -> int:
    mapping = {
        GovPublicSectorAIContext.BENEFITS_ELIGIBILITY_AI:   THRESHOLD_BENEFITS_ELIGIBILITY_AI,
        GovPublicSectorAIContext.LAW_ENFORCEMENT_INTEL_AI:  THRESHOLD_LAW_ENFORCEMENT_INTEL_AI,
        GovPublicSectorAIContext.PERMITTING_LICENSING_AI:   THRESHOLD_PERMITTING_LICENSING_AI,
        GovPublicSectorAIContext.VOTER_REGISTRATION_AI:     THRESHOLD_VOTER_REGISTRATION_AI,
    }
    return mapping[context]


async def scan_gov_public_sector_ai_image(
    image_path: str | Path,
    context: GovPublicSectorAIContext,
    agency_id_hash: str,           # SHA-256 of government agency or FIPS code identifier
    citizen_or_case_ref: str,      # e.g. "SNAP-CASE-2026-44821", "PERM-2026-88841"
    processing_session_id: str,    # document scan batch, intelligence analysis session ID
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan a government or public sector AI image for adversarial injection payloads
    before forwarding to benefits eligibility determination, law enforcement intelligence
    link analysis, permitting/licensing adjudication, or voter registration eligibility
    classification AI systems.

    Raises AdversarialGovPublicSectorAIImageError if score meets threshold:
      - BENEFITS_ELIGIBILITY_AI:   threshold 55; §1983; ADA §504; Goldberg; Mathews
      - LAW_ENFORCEMENT_INTEL_AI:  threshold 60; 4th Amendment; Brady; Giglio; §1519
      - PERMITTING_LICENSING_AI:   threshold 60; state APA; building code; OSHA; §1983
      - VOTER_REGISTRATION_AI:     threshold 65; NVRA; VRA; HAVA; §1983; §594
    """
    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": {
                "gov_public_sector_context": context.value,
                "agency_id_hash":            agency_id_hash,
                "citizen_or_case_ref":       citizen_or_case_ref,
                "processing_session_id":     processing_session_id,
                "client_scan_id":            client_scan_id,
                "image_sha256":              image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "agency_id_hash":           agency_id_hash,
        "citizen_or_case_ref":      citizen_or_case_ref,
        "processing_session_id":    processing_session_id,
        "gov_public_sector_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_gov_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialGovPublicSectorAIImageError(
            f"Government AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"agency={agency_id_hash} ref={citizen_or_case_ref}"
        )
    return result


async def write_gov_audit_record(record: dict) -> None:
    """Persist audit record to government compliance and civil rights documentation store (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialGovPublicSectorAIImageError(Exception):
    """Raised when a government or public sector AI image exceeds the adversarial injection threshold."""
    pass

Call scan_gov_public_sector_ai_image() with GovPublicSectorAIContext.BENEFITS_ELIGIBILITY_AI before forwarding Tyler Technologies AI or Salesforce Government Cloud benefits eligibility document photographs to eligibility indicator classification and programme qualification AI — with citizen_or_case_ref linking the Glyphward scan to the case record for 42 USC §1983 civil rights, ADA §504, and Goldberg v. Kelly pre-deprivation hearing compliance documentation. Call with GovPublicSectorAIContext.LAW_ENFORCEMENT_INTEL_AI for Palantir Gotham AI or Motorola Solutions CommandCentral AI intelligence link analysis visualisation displays before AI criminal indicator classification and investigative priority assessment, with agency_id_hash for Fourth Amendment exclusionary rule, Brady v. Maryland prosecutorial disclosure, and 18 USC §1519 obstruction audit trail documentation. Call with GovPublicSectorAIContext.PERMITTING_LICENSING_AI for Tyler Technologies AI or NIC AI permitting and licensing document scan images before AI compliance indicator classification and licence eligibility determination, with processing_session_id as the permit application batch identifier for state APA administrative record integrity and 42 USC §1983 property and liberty interest deprivation documentation. Call with GovPublicSectorAIContext.VOTER_REGISTRATION_AI for ES&S AI, Dominion Voting Systems AI, or state SOS voter registration AI document scan images before AI voter eligibility indicator classification and registration status determination, with agency_id_hash for NVRA voter registration rights, Voting Rights Act, HAVA provisional ballot, and 18 USC §594 voter intimidation compliance audit trail. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four government and public sector AI injection surfaces; covers how adversarial pixel-level perturbations cause AI misclassification without detectable visual artefacts at human review resolution.
• Vision-language model security — technical architecture of adversarial image attacks against vision-language models including pixel perturbation classes applicable to government document photograph AI, law enforcement intelligence visualisation AI, and voter registration document scan AI processing pipelines.
• Healthcare AI prompt injection — related regulatory framework covering HIPAA, ADA, and federal programme compliance applicable to government health benefit programme eligibility AI contexts including Medicaid and ACA marketplace AI eligibility determination tools.
• Free tier — 10 scans/day, no card required — start scanning government and public sector AI images at development volumes before committing to a production plan.