Prompt injection in HR and workforce AI

Four adversarial injection surfaces in HR and workforce AI

1. Video interview frame injection (HireVue AI, Modern Hire AI)

Video interview frame AI processes candidate video interview frame sequences from HireVue AI at more than 700 enterprise customers globally processing millions of candidate video interviews annually in financial services, technology, healthcare, and retail sector hiring programmes, Modern Hire AI at enterprise talent acquisition programme deployments, Pymetrics AI at Unilever, Goldman Sachs, and enterprise employer deployments, and integrated ATS platform video interview AI tools, extracting competency signal classifications — verbal communication effectiveness scores, non-verbal behavioural indicator assessments, job-relevant skill demonstration identifications, personality trait competency indicator ratings, and EEOC job-relatedness criterion competency dimension evaluations — from candidate video interview frame sequence inputs in AI-assisted hiring decision support pipelines, generating candidate competency score records, hiring manager review recommendations, interview stage advancement decisions, and candidate selection documentation entries that talent acquisition teams and EEOC compliance officers depend upon for EEOC Uniform Guidelines on Employee Selection Procedures job-related selection criterion compliance and NYC Local Law 144 automated employment decision tool bias audit obligation fulfilment. HireVue AI’s video interview AI processes candidate video interview frames through AI-assisted competency signal extraction and job-relatedness validated scoring tools that enterprise employers use for high-volume screening of entry-level and early-career hiring programmes where the volume of candidate applications makes individual human recruiter video review impracticable. NYC Local Law 144 (Automated Employment Decision Tools) requires employers who use AEDT in employment decisions affecting New York City employees to conduct independent bias audits, publish bias audit results, and notify affected candidates of AEDT use; HireVue AI video interview tools used in hiring processes for NYC employees constitute AEDT within the scope of Local Law 144 when used to screen or rank candidates for employment decisions. The Illinois AI Video Interview Act 820 ILCS 42 requires employers who use AI to analyse video interviews of Illinois applicants to notify applicants that AI may be used, explain how the AI works and what characteristics it evaluates, and obtain applicant consent before using AI video analysis in employment decisions.

The adversarial injection surface is the candidate video interview frame submission pathway: HireVue AI or Modern Hire AI candidate video interview frame sequences submitted through AI-assisted competency signal extraction, job-relevant skill indicator classification, and candidate quality scoring tools for AI hiring recommendation generation and candidate advancement decision support. An adversarially crafted HireVue AI video interview frame — in which pixel perturbations applied to the candidate verbal communication engagement visual indicator, the non-verbal body language signal region, or the job-relevant competency demonstration display in a candidate video interview frame sequence cause the AI to classify a candidate demonstrating strong job-relevant competency signals meeting the hiring recommendation threshold as falling below the competency signal threshold for positive hiring recommendation when the actual video interview frame documents candidate competency indicators meeting HireVue AI’s job-relatedness validated competency scoring criteria — can suppress a candidate competency score that would otherwise generate a positive hiring recommendation, an interview stage advancement decision, and a candidate selection pipeline progression record. Adversarial video interview frame injection that systematically suppresses competency scores for candidates from specific demographic groups creates EEOC Title VII disparate impact exposure — where an employment selection procedure that produces different passage rates for members of a protected class in a statistically significant adverse impact ratio constitutes prima facie evidence of disparate impact discrimination under the 80% rule of EEOC Uniform Guidelines on Employee Selection Procedures 29 CFR Part 1607, shifting the burden to the employer to demonstrate job-relatedness and business necessity for the selection procedure.

The EEOC and NYC Local Law 144 consequences of adversarially manipulated competency classification in video interview AI span Title VII disparate impact, ADA reasonable accommodation, NYC Local Law 144 AEDT bias audit, and Illinois AI Video Interview Act dimensions. EEOC Title VII of the Civil Rights Act of 1964 prohibits employment practices that have a disparate impact on a protected class — race, colour, religion, sex, or national origin — unless the practice is job-related and consistent with business necessity; adversarial manipulation of HireVue AI video interview competency scoring that systematically suppresses scores for protected class candidates creates the disparate impact statistical signature (adverse impact ratio below 4/5 rule threshold) that triggers EEOC investigation and shifts the job-relatedness burden of proof to the employer. ADA Title I prohibits employment discrimination against qualified individuals with disabilities; HireVue AI video interview AI that uses non-verbal behavioural indicators in competency assessment creates ADA disparate impact dimensions when adversarial manipulation of video frame AI suppresses competency scores for candidates whose non-verbal communication patterns reflect disability-related characteristics, creating an AI-generated selection barrier not justified by job-related criteria. NYC Local Law 144 requires independent AEDT bias audits by independent auditors using the EEOC Uniform Guidelines 4/5 rule or other statistically validated adverse impact analysis methodology; adversarially manipulated HireVue AI video interview competency scoring that inflates adverse impact ratios creates Local Law 144 bias audit non-compliance and  $375-$1,500 per day per violation civil penalty exposure. Threshold: 60 for video interview frame AI — reflecting the EEOC Title VII disparate impact, ADA employment non-discrimination, NYC Local Law 144, and Illinois AI Video Interview Act dimensions of adversarially manipulated competency classification.

2. Resume document scan injection (Paradox Olivia AI, Eightfold AI)

Resume document scan AI processes candidate resume and job application document scan images from Paradox Olivia AI across more than 100 million ATS-integrated candidate conversations at enterprise talent acquisition programmes globally, Eightfold AI talent intelligence platform at enterprise talent acquisition and internal mobility programme deployments, Workday AI ATS platform candidate resume processing tools at more than 10,000 enterprise customers, SAP SuccessFactors AI talent management platform resume screening tools, and integrated ATS platform AI resume analysis tools, extracting candidate qualification and job-relevancy classifications — skills and competency keyword match scores, career trajectory progression assessments, education credential relevancy ratings, employment history continuity evaluations, and EEOC job-related qualification criterion match determinations — from candidate resume document scan image inputs in AI-assisted applicant screening and candidate pool management pipelines, generating candidate qualification scores, applicant screening pass or fail decisions, candidate pool inclusion or exclusion determinations, and ATS applicant ranking records that recruiters and hiring managers depend upon for EEOC-compliant applicant screening and OFCCP federal contractor applicant flow data documentation. Paradox Olivia AI’s resume screening AI processes candidate application documents through AI-assisted job-relevancy qualification screening tools that ATS-integrated enterprise employers use for high-volume entry-level and professional role candidate screening at organisations including Unilever, McDonald’s, and major financial services employers where application volumes make individual human recruiter resume review of every applicant impracticable. Eightfold AI’s talent intelligence platform processes candidate career trajectory data through AI-assisted skills graph mapping and candidate-to-role matching tools that talent acquisition teams use for both external candidate pipeline screening and internal employee mobility matching with EEOC adverse impact monitoring dimensions.

The adversarial injection surface is the candidate resume and job application document scan image submission pathway: Paradox Olivia AI or Eightfold AI candidate resume document scan images submitted through AI-assisted skills matching, career trajectory assessment, and job-relevancy qualification scoring tools for AI candidate qualification classification and ATS applicant screening decision. An adversarially crafted Eightfold AI resume document scan image — in which pixel perturbations applied to the candidate skills keyword display region, the employment history continuity visual indicator, or the educational credential relevancy marker in a resume document scan image cause the AI to classify a qualified candidate meeting the job-relevant qualification criteria for ATS screening advancement as a below-threshold unqualified applicant not meeting the minimum qualification criteria when the actual resume document displays qualifications meeting Eightfold AI’s job-relevancy scoring threshold for candidate pool inclusion — can suppress a candidate qualification score that would otherwise generate an ATS screening pass decision, a recruiter review recommendation, and a candidate pool inclusion record. Adversarial resume scan injection that systematically suppresses qualification scores for candidates whose resume document images contain visual features statistically correlated with protected class membership — such as historically Black college and university name visual markers, women’s leadership programme certification logos, or non-anglophone name visual patterns — creates EEOC Title VII disparate impact exposure with OFCCP applicant flow data integrity and federal contractor adverse impact analysis obligation dimensions.

The EEOC and OFCCP consequences of adversarially manipulated qualification classification in resume document scan AI span Title VII disparate impact, OFCCP federal contractor adverse impact analysis, ADA disability discrimination, and state fair employment practices statute dimensions. EEOC Uniform Guidelines on Employee Selection Procedures 29 CFR Part 1607 specifies that any employment selection procedure that has an adverse impact — defined as a selection rate for a protected group less than four-fifths of the highest selection rate for any group — is presumptively discriminatory unless the employer demonstrates job-relatedness and business necessity; adversarial manipulation of Paradox Olivia AI or Eightfold AI resume screening that systematically suppresses qualification scores for protected class candidates creates the EEOC adverse impact signature that triggers EEOC investigation and Charge of Discrimination filing authority. OFCCP 41 CFR Part 60-2 requires federal contractors to conduct annual workforce and compensation analyses including applicant flow data adverse impact analysis for each job group; adversarial manipulation of ATS resume screening AI that corrupts the applicant flow data by excluding qualified protected class candidates from candidate pool records creates OFCCP audit data integrity failures that affect the contractor’s compliance certification and annual affirmative action programme accuracy. California Fair Employment and Housing Act (FEHA) Government Code §12940, New York Human Rights Law Executive Law §296, and Illinois Human Rights Act 775 ILCS 5/2-102 prohibit employment discrimination based on protected characteristics in addition to federal Title VII requirements; adversarial manipulation of resume AI that creates state-law protected class disparate impact in California, New York, or Illinois creates parallel state law employment discrimination liability exposure with broader protected class categories than federal Title VII. Threshold: 60 for resume document scan AI — reflecting the EEOC Title VII disparate impact, OFCCP adverse impact analysis integrity, ADA employment non-discrimination, and state fair employment practices dimensions of adversarially manipulated qualification classification.

3. Compensation benchmark data visualisation injection (Beqom AI, Visier AI)

Compensation benchmark data visualisation AI processes compensation market benchmark data display graphics, pay equity gap analysis visualisation displays, salary band positioning data graphics, and compensation equity audit report images from Beqom AI at enterprise HR compensation management programme deployments globally, Visier AI people analytics platform at enterprise HR analytics programme deployments including large financial services, technology, and healthcare sector employers, Workday AI compensation management module analytics displays, and integrated HRIS platform compensation equity analysis AI tools, extracting compensation equity and market positioning classifications — gender pay gap magnitude assessments, racial pay equity gap identifications, job group compensation band positioning determinations, and OFCCP proactive pay equity audit compliance indicators — from compensation benchmark visualisation display image inputs in AI-assisted compensation equity analysis and OFCCP compliance monitoring pipelines, generating pay equity remediation priority alerts, compensation adjustment recommendation records, OFCCP proactive compliance certification inputs, and state pay equity statute compliance documentation entries that total rewards and HR compliance teams depend upon for Equal Pay Act 29 USC §206(d) compliance monitoring and state pay equity statute audit obligation management. Beqom AI’s compensation analytics platform processes compensation benchmark data visualisations and pay equity gap analysis displays through AI-assisted pay equity gap identification and OFCCP proactive compliance analysis tools that enterprise total rewards teams use for Equal Pay Act and state pay equity statute compliance monitoring at organisations with pay equity audit obligations under California Labor Code §1197.5, Colorado Equal Pay for Equal Work Act C.R.S. §8-5-101, and Washington Equal Pay and Opportunities Act RCW 49.58. Visier AI’s people analytics platform processes workforce analytics dashboard displays and compensation equity data visualisations through AI-assisted pay equity insight and headcount equity classification tools at enterprise HR analytics programmes where AI-generated pay equity dashboards inform compensation review cycle decisions and OFCCP compliance programme documentation.

The adversarial injection surface is the compensation benchmark data visualisation display and pay equity gap analysis graphic submission pathway: Beqom AI or Visier AI compensation equity data visualisation display images submitted through AI-assisted pay equity gap magnitude assessment, gender and racial pay disparity identification, and OFCCP proactive compliance analysis tools for AI pay equity classification and compensation remediation priority determination. An adversarially crafted Beqom AI compensation benchmark data visualisation display image — in which pixel perturbations applied to the gender pay gap magnitude display indicator, the job group compensation band positioning visual marker, or the racial pay equity gap data bar chart region in a compensation equity analysis visualisation cause the AI to classify a compensation equity dataset with statistically significant gender or racial pay equity gaps meeting OFCCP proactive compliance audit threshold criteria as a compensation equity profile below the pay equity remediation trigger threshold when the actual visualisation documents gender or racial pay disparities requiring compensation adjustment under Equal Pay Act 29 USC §206(d) and OFCCP federal contractor proactive compensation equity compliance — can suppress a pay equity remediation priority alert that would otherwise generate a compensation adjustment recommendation, an OFCCP proactive compliance certification entry, and a pay equity audit documentation record. In enterprise compensation management environments where Beqom AI or Visier AI processes compensation equity data visualisations for large employer workforces without individual total rewards analyst review of every AI pay equity classification, adversarial suppression of gender or racial pay gap identifications allows pay equity violations to go undetected and unremediated in compensation review cycles with Equal Pay Act, OFCCP, and state pay equity statute compliance consequences.

The Equal Pay Act and state pay equity statute consequences of adversarially suppressed pay gap identification in compensation analytics AI span Equal Pay Act 29 USC §206(d), OFCCP proactive compensation equity compliance, California Labor Code §1197.5, Colorado C.R.S. §8-5-101, Washington RCW 49.58, and New York Equal Pay for Equal Work Labor Law §194 dimensions. Equal Pay Act 29 USC §206(d) prohibits wage differentials between male and female employees based on sex who perform substantially equal work in the same establishment; adversarial manipulation of Beqom AI or Visier AI compensation analytics that suppresses gender pay gap identification allows Equal Pay Act violations to persist in compensation structures without generating the remediation alerts that total rewards teams use to correct pay disparities before they accumulate to the magnitude that supports collective action wage and hour litigation. OFCCP Directive 2022-01 specifies that OFCCP will conduct proactive supply and service contractor compensation equity evaluations using regression-based statistical analysis to identify pay disparities by race and gender; adversarially manipulated Visier AI or Beqom AI compensation analytics dashboard displays that suppress pay gap identification create OFCCP compliance audit documentation failures when the contractor’s OFCCP proactive compliance certification does not reflect actual compensation disparities identified in workforce compensation data. California Labor Code §1197.5 (as amended by SB 1162 effective January 2023) requires employers with 100 or more employees to submit annual pay data reports to the California Civil Rights Department covering median and mean hourly rates by race, ethnicity, and sex for each job category; adversarially suppressed Beqom AI or Visier AI pay equity analytics that produces inaccurate pay data report inputs creates California pay data reporting accuracy obligations and California Civil Rights Department enforcement exposure. Threshold: 65 for compensation benchmark data visualisation AI — reflecting the Equal Pay Act, OFCCP proactive compliance audit, California SB 1162 pay data reporting, and state pay equity statute dimensions of adversarially suppressed pay gap identification.

4. Workforce analytics dashboard injection (Workday AI, Visier AI)

Workforce analytics dashboard AI processes workforce headcount, attrition, and performance distribution dashboard display images from Workday AI at more than 10,000 enterprise customers including major federal government contractors, Visier AI people analytics platform at enterprise HR analytics programme deployments, UiPath HR process automation AI at HR operations workflow and analytics automation deployments, SAP SuccessFactors AI HCM suite analytics displays, and integrated HRIS platform workforce analytics AI tools, extracting workforce disparity and compliance risk classifications — headcount equity demographic representation gap indicators, attrition disparity by protected class rate detections, performance rating distribution adverse impact flags, promotion rate equity disparity identifications, and OFCCP adverse impact analysis trigger determinations — from workforce analytics dashboard display image inputs in AI-assisted HR compliance monitoring and EEOC reporting pipelines, generating EEOC EEO-1 Component 1 reporting accuracy verifications, OFCCP adverse impact analysis trigger notifications, workforce equity remediation priority alerts, and HR compliance programme documentation entries that HR compliance and D&I leadership teams depend upon for EEOC EEO-1 reporting obligation compliance and OFCCP federal contractor annual affirmative action programme accuracy. Workday AI’s workforce analytics tools process headcount and demographic distribution dashboard visualisations through AI-assisted EEOC EEO-1 data accuracy verification and OFCCP adverse impact analysis trigger identification tools at large enterprise customers where AI-generated workforce analytics dashboards inform EEOC EEO-1 Component 1 annual reporting and OFCCP affirmative action programme development. Visier AI’s people analytics platform processes workforce performance and attrition distribution dashboards through AI-assisted equity insight and disparity classification tools at enterprise HR programmes where AI-generated workforce equity dashboards inform D&I programme investment decisions and OFCCP compliance monitoring.

The adversarial injection surface is the workforce analytics dashboard display and EEOC compliance reporting data visualisation submission pathway: Workday AI or Visier AI workforce analytics dashboard display images submitted through AI-assisted headcount equity assessment, attrition and performance distribution adverse impact analysis, and OFCCP compliance trigger identification tools for AI workforce disparity classification and HR compliance programme priority determination. An adversarially crafted Workday AI workforce analytics dashboard display image — in which pixel perturbations applied to the demographic headcount representation gap bar chart region, the protected class attrition rate disparity indicator display, or the performance distribution adverse impact ratio visual marker in a workforce analytics dashboard image cause the AI to classify a workforce analytics dataset exhibiting statistically significant protected class representation gaps or performance distribution disparities meeting OFCCP adverse impact analysis trigger thresholds as a workforce equity profile below the compliance intervention trigger threshold when the actual dashboard documents workforce disparities requiring OFCCP AAP remediation programme development — can suppress a workforce disparity alert that would otherwise generate an OFCCP adverse impact analysis notification, a headcount equity remediation priority record, and an affirmative action programme goal-setting documentation entry. In large enterprise workforce analytics environments where Workday AI or Visier AI processes workforce dashboard visualisations across large federal contractor workforce populations without individual HR compliance analyst review of every AI disparity classification, adversarial suppression of demographic representation gap detections allows EEOC and OFCCP reportable workforce disparities to go unaddressed in affirmative action programme development with OFCCP conciliation agreement and debarment risk consequences.

The EEOC and OFCCP consequences of adversarially suppressed workforce disparity detection in workforce analytics dashboard AI span EEOC EEO-1 reporting accuracy, OFCCP affirmative action programme compliance, EEOC Pattern or Practice investigation trigger, and federal contractor debarment dimensions. EEOC EEO-1 Component 1 reporting requires covered employers with 100 or more employees to file annual demographic workforce data reports with the EEOC; EEOC uses EEO-1 data in systemic discrimination pattern or practice investigations and charges. Adversarially suppressed Workday AI workforce analytics that produces inaccurate EEOC EEO-1 reporting data creates EEOC EEO-1 data accuracy obligation exposure and EEOC systemic investigation trigger consequence when inaccurate EEO-1 data masks workforce disparities that would otherwise trigger EEOC pattern or practice investigation. OFCCP federal contractor compliance obligation under Executive Order 11246 and 41 CFR Part 60 requires federal contractors and subcontractors to develop annual affirmative action programmes including availability analysis, utilisation analysis, and goal-setting for underutilised protected class groups; adversarial manipulation of Workday AI workforce analytics dashboard AI that suppresses protected class utilisation gap detections creates OFCCP affirmative action programme goal-setting accuracy failures that OFCCP compliance review auditors assess in scheduled compliance evaluations. OFCCP debarment of federal contractors under 41 CFR Part 60-1.26 for failure to comply with Equal Employment Opportunity and affirmative action requirements is the most severe OFCCP enforcement consequence; adversarially manipulated workforce analytics AI that prevents detection of OFCCP-reportable workforce disparities and creates affirmative action programme compliance failures creates debarment risk dimensions for federal contractors who rely on OFCCP compliance programme documentation generated by adversarially corrupted workforce analytics AI. Threshold: 65 for workforce analytics dashboard AI — reflecting the EEOC EEO-1 reporting accuracy, OFCCP adverse impact analysis, affirmative action programme goal-setting integrity, and federal contractor debarment dimensions of adversarially suppressed workforce disparity detection.

Integration: HR and workforce AI image ingestion with Glyphward pre-scan

HR and workforce AI image ingestion flows from HireVue and Modern Hire candidate video interview frame APIs, Paradox Olivia and Eightfold resume document scan image channels, Beqom and Visier compensation benchmark data visualisation interfaces, and Workday and Visier workforce analytics dashboard display platforms into competency signal extraction and hiring decision support AI, job-relevancy qualification screening and OFCCP applicant flow AI, pay equity gap identification and Equal Pay Act compliance AI, and workforce disparity detection and EEOC EEO-1 reporting AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary before AI-generated output is committed to candidate hiring recommendations, resume screening pass/fail decisions, pay equity remediation priorities, or OFCCP adverse impact analysis triggers:

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"

# HR & workforce AI — EEOC Title VII 42 USC §2000e-2 disparate impact;
# ADA Title I 42 USC §12112; NYC Local Law 144 AEDT; Illinois AI Video
# Interview Act 820 ILCS 42; OFCCP 41 CFR Part 60; Equal Pay Act
# 29 USC §206(d); CA Labor Code §1197.5 SB 1162; EEOC EEO-1 Component 1.
THRESHOLD_VIDEO_INTERVIEW_AI   = 60  # HireVue/Modern Hire; Title VII; Local Law 144
THRESHOLD_RESUME_SCAN_AI       = 60  # Paradox/Eightfold; OFCCP applicant flow; ADA
THRESHOLD_COMPENSATION_AI      = 65  # Beqom/Visier; Equal Pay Act; CA SB 1162
THRESHOLD_WORKFORCE_ANALYTICS_AI = 65  # Workday/Visier; EEOC EEO-1; OFCCP debarment


class HRWorkforceAIContext(str, Enum):
    VIDEO_INTERVIEW_AI      = "video_interview_ai"      # HireVue, Modern Hire
    RESUME_SCAN_AI          = "resume_scan_ai"          # Paradox, Eightfold
    COMPENSATION_AI         = "compensation_ai"         # Beqom, Visier pay equity
    WORKFORCE_ANALYTICS_AI  = "workforce_analytics_ai" # Workday, Visier analytics


def threshold_for(context: HRWorkforceAIContext) -> int:
    mapping = {
        HRWorkforceAIContext.VIDEO_INTERVIEW_AI:      THRESHOLD_VIDEO_INTERVIEW_AI,
        HRWorkforceAIContext.RESUME_SCAN_AI:          THRESHOLD_RESUME_SCAN_AI,
        HRWorkforceAIContext.COMPENSATION_AI:         THRESHOLD_COMPENSATION_AI,
        HRWorkforceAIContext.WORKFORCE_ANALYTICS_AI:  THRESHOLD_WORKFORCE_ANALYTICS_AI,
    }
    return mapping[context]


async def scan_hr_workforce_ai_image(
    image_path: str | Path,
    context: HRWorkforceAIContext,
    employer_id_hash: str,       # SHA-256 of employer or OFCCP contractor identifier
    candidate_or_report_ref: str, # e.g. "HRV-CAND-2026-44821", "BEQOM-RPT-88841"
    evaluation_session_id: str,  # interview session, resume scan run, or report period
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan an HR or workforce AI image for adversarial injection payloads before
    forwarding to video interview competency classification, resume screening
    qualification scoring, compensation equity gap identification, or workforce
    disparity and EEOC EEO-1 analytics AI systems.

    Raises AdversarialHRWorkforceAIImageError if score meets threshold:
      - VIDEO_INTERVIEW_AI:      threshold 60; Title VII; ADA; NYC LL 144; IAVIA
      - RESUME_SCAN_AI:          threshold 60; Title VII; OFCCP applicant flow; ADA
      - COMPENSATION_AI:         threshold 65; Equal Pay Act; OFCCP; CA SB 1162
      - WORKFORCE_ANALYTICS_AI:  threshold 65; EEOC EEO-1; OFCCP debarment risk
    """
    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": {
                "hr_workforce_context":       context.value,
                "employer_id_hash":           employer_id_hash,
                "candidate_or_report_ref":    candidate_or_report_ref,
                "evaluation_session_id":      evaluation_session_id,
                "client_scan_id":             client_scan_id,
                "image_sha256":               image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "employer_id_hash":          employer_id_hash,
        "candidate_or_report_ref":   candidate_or_report_ref,
        "evaluation_session_id":     evaluation_session_id,
        "hr_workforce_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_hr_workforce_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialHRWorkforceAIImageError(
            f"HR/workforce AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"employer={employer_id_hash} ref={candidate_or_report_ref}"
        )
    return result


async def write_hr_workforce_audit_record(record: dict) -> None:
    """Persist audit record to HR compliance and EEOC/OFCCP documentation store (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialHRWorkforceAIImageError(Exception):
    """Raised when an HR or workforce AI image exceeds the adversarial injection threshold."""
    pass

Call scan_hr_workforce_ai_image() with HRWorkforceAIContext.VIDEO_INTERVIEW_AI before forwarding HireVue AI or Modern Hire candidate video interview frames to competency signal classification and hiring recommendation AI — with candidate_or_report_ref linking the Glyphward scan to the candidate record for EEOC Title VII adverse impact analysis, NYC Local Law 144 AEDT bias audit, and Illinois AI Video Interview Act compliance documentation. Call with HRWorkforceAIContext.RESUME_SCAN_AI for Paradox Olivia or Eightfold resume document scan images before AI qualification scoring and ATS screening decisions, with employer_id_hash for OFCCP federal contractor applicant flow data integrity and adverse impact analysis audit trail. Call with HRWorkforceAIContext.COMPENSATION_AI for Beqom or Visier compensation equity data visualisation displays before AI pay equity gap identification, with evaluation_session_id as the compensation review period identifier for Equal Pay Act, OFCCP proactive compliance, and California SB 1162 pay data reporting audit documentation. Call with HRWorkforceAIContext.WORKFORCE_ANALYTICS_AI for Workday or Visier workforce analytics dashboard displays before AI headcount equity and EEOC EEO-1 disparity classification, with employer_id_hash for EEOC EEO-1 Component 1 reporting accuracy and OFCCP affirmative action programme goal-setting documentation audit trail. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four HR and workforce 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 video interview frame AI and resume document scan AI processing pipelines.
• Financial services AI prompt injection — related regulatory framework covering CFPB, ECOA fair lending, and financial AI integrity applicable to Workday AI and SAP SuccessFactors AI compensation analytics contexts.
• Free tier — 10 scans/day, no card required — start scanning HR and workforce AI images at development volumes before committing to a production plan.