Prompt injection in mental health and digital health AI

Four adversarial injection surfaces in mental health and digital health AI

1. Mental health assessment photograph injection (Spring Health AI, Lyra Health AI)

Mental health assessment photograph AI processes patient affect and mental state self-assessment photographs from Spring Health AI at more than 1,000 employer customers covering millions of employees globally, Lyra Health AI employer benefit programme deployments at global enterprise customers across financial services, technology, and healthcare sectors, Ginger AI (Headspace Health) employer benefit and health plan coverage programme deployments, and SilverCloud Health AI (Amwell) digital therapeutic platform deployments, extracting depression and mental health severity classifications — PHQ-9 symptom severity score indicators, anxiety severity assessment ratings, emotional distress presentation markers, clinical deterioration risk flags — from patient-submitted affect and mental state photograph inputs in AI-assisted mental health assessment pipelines, generating clinical care pathway recommendations, therapist escalation triggers, urgent outreach notifications, and MHPAEA-compliant clinical necessity determinations that employer benefit programme administrators and clinician care teams depend upon for timely mental health crisis intervention and HIPAA-compliant patient care management. Spring Health AI’s assessment photograph AI is a primary AI-assisted severity classification mechanism for employer benefit programme mental health care navigation at scale; its patient assessment photograph AI processes PHQ-9 and GAD-7 equivalent symptom severity assessment inputs through AI-assisted severity classification tools that determine whether a patient’s current symptom presentation meets the threshold for urgent clinical outreach, crisis intervention referral, or standard care pathway assignment under MHPAEA-compliant clinical necessity criteria and employer benefit programme mental health care escalation protocols. Joint Commission NPSG.15.01.01 (Suicide Risk Reduction) specifies that accredited healthcare organisations implement evidence-based processes for identifying patients at risk for suicide; AI-assisted mental health assessment photograph severity classification that feeds clinical escalation decisions in Joint Commission-accredited employer health programme contexts creates NPSG.15.01.01 compliance dimensions when AI severity output governs whether patients presenting with PHQ-9 severe depression indicators receive timely crisis intervention.

The adversarial injection surface is the patient mental health assessment photograph submission pathway: Spring Health AI or Lyra Health AI patient affect and mental state assessment photographs submitted through AI-assisted PHQ-9 severity classification, emotional distress presentation assessment, and clinical care pathway escalation tools for AI depression severity identification and clinical intervention trigger determination. An adversarially crafted Spring Health AI mental health assessment photograph — in which pixel perturbations applied to the patient affect presentation visual markers, the emotional distress expression region, or the clinical deterioration indicator display in a patient self-assessment photograph cause the AI to classify a patient presenting with PHQ-9 severe depression symptom indicators as meeting only mild symptom severity criteria falling below the clinical escalation threshold when the actual photograph documents a patient affect presentation meeting PHQ-9 severe depression criteria requiring urgent clinical outreach — can suppress a clinical escalation trigger that would otherwise generate an urgent care team notification, a crisis intervention referral, and a MHPAEA-compliant clinical necessity documentation record. In employer benefit programme deployments where Spring Health AI or Lyra Health AI processes thousands of patient self-assessment photograph submissions per day across large enterprise employer benefit populations without individual clinician review of each AI severity classification, adversarial suppression of PHQ-9 severe depression severity classifications allows patients in clinical crisis to be routed to standard care pathways without the urgent clinical intervention that their actual presentation requires, with Tarasoff duty-to-warn and Joint Commission NPSG.15.01.01 consequences when the suppressed crisis assessment delays clinical intervention.

The regulatory and duty-to-warn consequences of adversarially suppressed mental health severity classification in assessment photograph AI span HIPAA, APA Ethics Code, Tarasoff common law, MHPAEA, and Joint Commission NPSG dimensions. HIPAA 45 CFR §164.530(c) requires covered entities to implement policies and procedures that protect PHI and ensure the accuracy of clinical information used for treatment decisions; adversarially manipulated mental health assessment photograph AI that generates inaccurate PHQ-9 severity classifications used in patient care pathway determination creates HIPAA §164.530(c) clinical information accuracy obligation exposure when the adversarially corrupted assessment record becomes part of the patient’s HIPAA-protected treatment record. APA Ethics Code 4.05(b) permits psychologists to disclose confidential information to protect an identified third party from a serious and imminent threat; Tarasoff v. Regents of the University of California (1976) 17 Cal. 3d 425 established a common law duty for mental health clinicians to warn identified potential victims when a patient presents a serious and credible threat of harm. Adversarial suppression of Spring Health AI or Lyra Health AI mental health severity classification that prevents the AI-assisted assessment tool from generating the crisis severity indicator that would trigger a clinician Tarasoff review creates a duty-to-warn liability exposure when the suppressed severity classification delays clinician awareness of a patient risk presentation that, if accurately classified, would have triggered a duty-to-warn evaluation. Threshold: 50 for mental health assessment photograph AI — reflecting the Tarasoff duty-to-warn liability, HIPAA PHI clinical accuracy, MHPAEA clinical necessity, and Joint Commission NPSG.15.01.01 suicide risk reduction dimensions of suppressed mental health severity classification.

2. Crisis support interface screenshot injection (Crisis Text Line AI, Woebot AI)

Crisis support interface screenshot AI processes crisis text support conversation interface screenshots from Crisis Text Line AI at national crisis intervention service operations providing 24/7 crisis text-based support under the 988 Suicide and Crisis Lifeline national network, Woebot Health AI conversational AI mental health and crisis support platform at consumer and employer benefit deployments, Ginger AI (Headspace Health) crisis escalation tool at employer benefit and health plan programme deployments, and integrated digital mental health platform crisis support AI tools, extracting crisis severity classifications — suicidal ideation risk scores, self-harm intent probability assessments, crisis severity level determinations, emergency response protocol trigger indicators — from crisis support conversation interface screenshot inputs in AI-assisted crisis triage and escalation pipelines, generating crisis counsellor intervention alerts, 988 Lifeline emergency response escalation triggers, emergency services dispatch notifications, and joint crisis intervention team coordination records that crisis intervention organisations and clinical teams depend upon for timely suicidal crisis intervention and Joint Commission NPSG.15.01.01 suicide risk reduction standard compliance. Crisis Text Line AI’s crisis support AI processes more than 6 million crisis text conversations through AI-assisted crisis severity classification and counsellor matching tools that determine which conversations receive priority counsellor assignment, crisis severity escalation, and emergency services notification under 988 Lifeline response protocols and Joint Commission NPSG.15.01.01 suicide risk assessment requirements. The 988 Suicide and Crisis Lifeline national network established by SAMHSA under 34 USC §290bb-36b specifies crisis intervention response protocols that include suicidal ideation risk triage, active rescue protocol implementation, and emergency services coordination; AI-assisted crisis support interface screenshot severity classification tools that feed 988 Lifeline counsellor triage and emergency response escalation decisions operate within the federal crisis intervention framework where AI output accuracy directly affects whether individuals in suicidal crisis receive timely emergency response.

The adversarial injection surface is the crisis support conversation interface screenshot submission pathway: Crisis Text Line AI or Woebot AI crisis support conversation session interface screenshots submitted through AI-assisted suicidal ideation risk classification, crisis severity level determination, and 988 Lifeline emergency response protocol trigger tools for AI crisis severity identification and crisis counsellor intervention assignment. An adversarially crafted Crisis Text Line AI crisis support interface screenshot — in which pixel perturbations applied to the suicidal ideation risk indicator display region, the crisis severity level marker, or the emergency response protocol trigger condition visual in a crisis support conversation interface screenshot cause the AI to classify a crisis conversation exhibiting suicidal ideation risk indicators meeting the 988 Lifeline active rescue protocol threshold as a standard-severity conversation meeting only routine counsellor support criteria when the actual screenshot documents crisis content meeting Joint Commission NPSG.15.01.01 high-risk classification criteria — can suppress a crisis severity escalation trigger that would otherwise generate a priority counsellor assignment, an emergency services dispatch notification, and a 988 Lifeline active rescue protocol implementation record. In high-volume crisis text support operations where Crisis Text Line AI processes thousands of concurrent crisis conversations per day across the national 988 Lifeline network without individual counsellor review of every AI severity triage result before priority assignment decisions, adversarial suppression of suicidal ideation risk severity classifications in crisis conversation screenshots allows individuals in acute suicidal crisis to be routed to standard counsellor queues without the priority crisis intervention response that their actual risk presentation requires, with catastrophic patient safety and 988 Lifeline protocol compliance consequences.

The patient safety and regulatory consequences of adversarially suppressed crisis severity classification in crisis support interface screenshot AI span Joint Commission NPSG.15.01.01, 988 Lifeline federal network response protocol, SAMHSA emergency services coordination, and Tarasoff common law duty-to-warn dimensions. Joint Commission NPSG.15.01.01 requires accredited hospitals and behavioural health organisations to implement evidence-based suicide risk assessment processes using validated suicide risk assessment tools; Crisis Text Line and integrated employer mental health benefit programme crisis support AI tools deployed in Joint Commission-accredited programme contexts create NPSG.15.01.01 compliance dimensions when AI-assisted crisis severity classification governs whether patients assessed through accredited programme crisis support interfaces receive timely suicide risk response. SAMHSA 988 Lifeline network participation requirements specify that network member crisis centre operations maintain crisis response capacity standards including suicidal ideation risk assessment and active rescue protocol implementation capability; adversarial manipulation of Crisis Text Line AI that suppresses suicidal ideation risk severity classifications in crisis conversation screenshots creates 988 Lifeline network participation standard compliance exposure when manipulated AI triage results systematically reduce active rescue protocol implementation rates. 42 USC §290bb-36b federal crisis intervention programme requirements specify federal funding compliance obligations for 988 Lifeline network member crisis centres; adversarially compromised crisis AI triage that reduces emergency response effectiveness creates federal programme participation obligation exposure with SAMHSA funding compliance dimensions. Threshold: 50 for crisis support interface screenshot AI — reflecting the patient life safety, Joint Commission NPSG.15.01.01, 988 Lifeline emergency response protocol, and Tarasoff duty-to-warn dimensions of suppressed crisis severity classification.

3. Teletherapy session video frame injection (Talkspace AI, BetterHelp AI)

Teletherapy session video frame AI processes therapist-patient teletherapy video session interface images and session monitoring frames from Talkspace AI at consumer and employer benefit programme teletherapy deployments covering millions of patients across the United States and international markets, BetterHelp AI at consumer teletherapy deployments representing the world’s largest online therapy platform with more than 3 million patients, SilverCloud Health AI (Amwell) digital therapeutic and teletherapy support programme tools, and integrated telehealth platform teletherapy AI session support tools, extracting session quality and clinical engagement classifications — therapeutic alliance indicator assessments, patient engagement and distress signal detections, session crisis emergence risk flags, and clinical note-assist content accuracy verifications — from teletherapy video session interface image inputs in AI-assisted teletherapy quality management and clinical support pipelines, generating therapist clinical alert notifications, supervisor escalation flags, session quality review triggers, and HIPAA-compliant clinical note documentation records that teletherapy platform clinical quality teams and therapist supervisors depend upon for AI-assisted teletherapy quality management and patient safety monitoring. Talkspace AI’s session monitoring AI processes teletherapy session quality indicators through AI-assisted therapeutic engagement and patient distress signal classification tools that generate therapist clinical support alerts and clinical supervisor escalation flags for sessions where AI-detected signals indicate patient safety risk or therapeutic quality concerns requiring clinical review. BetterHelp AI’s session management platform processes patient intake and teletherapy session matching interface images through AI-assisted clinical matching and session support tools at the world’s largest consumer teletherapy platform, where AI-assisted session quality monitoring and clinical note documentation support tools generate the clinical records that HIPAA-covered teletherapy providers maintain as treatment records with APA Ethics Code professional responsibility dimensions.

The adversarial injection surface is the teletherapy session video interface frame submission pathway: Talkspace AI or BetterHelp AI teletherapy video session interface screenshots and session monitoring frames submitted through AI-assisted patient engagement classification, crisis emergence detection, and clinical note documentation tools for AI therapeutic alliance assessment and patient safety signal identification. An adversarially crafted Talkspace AI teletherapy session video frame — in which pixel perturbations applied to the patient distress signal visual marker, the therapeutic engagement level indicator region, or the crisis emergence risk indicator display in a teletherapy session video interface frame cause the AI to classify a session exhibiting patient distress and crisis emergence signals meeting the clinical escalation threshold as a standard engagement session below the clinical alert threshold when the actual session frame documents patient distress indicators requiring therapist supervisor notification — can suppress a clinical alert that would otherwise generate a therapist supervisor escalation notification and a HIPAA-compliant clinical record entry documenting the patient distress signal and clinical response. In high-volume teletherapy session management environments where Talkspace AI or BetterHelp AI processes thousands of concurrent teletherapy session monitoring frames without individual clinical supervisor review of every AI session quality classification, adversarial suppression of patient distress signal detections and crisis emergence risk flags allows sessions with emerging patient safety risks to proceed without the clinical supervisor escalation that the patient’s actual session presentation requires.

The HIPAA and APA Ethics Code consequences of adversarially suppressed patient distress signal detection in teletherapy session video frame AI span HIPAA treatment record accuracy, APA Ethics Code professional supervision, Tarasoff duty-to-warn, and state telehealth practice standard dimensions. HIPAA 45 CFR §164.526 provides patients the right to amend PHI they believe to be inaccurate or incomplete; teletherapy session AI clinical note records that fail to document patient distress signals and clinical escalation events because adversarial image manipulation suppressed the AI detection that would have generated clinical note entries create HIPAA §164.526 PHI accuracy obligations when the adversarially incomplete session record is later accessed by the patient or a subsequent treating clinician. APA Ethics Code 2.05 requires licensed psychologists who delegate professional responsibilities to ensure that delegated functions are performed competently; teletherapy platform AI session monitoring tools that clinician supervisors rely upon for quality oversight of supervised therapist sessions create APA professional supervision responsibility dimensions when adversarial manipulation of session AI suppresses patient distress signals that responsible clinical supervision should have detected. State telehealth practice standards established by state licensing boards — including state-specific telehealth regulations for psychologists, licensed clinical social workers, and licensed professional counsellors — specify documentation standards for teletherapy session clinical records; adversarial suppression of Talkspace or BetterHelp AI session monitoring outputs that creates incomplete clinical documentation creates state telehealth practice standard compliance exposure with professional licensing consequences. Threshold: 50 for teletherapy session video frame AI — reflecting the HIPAA treatment record accuracy, APA Ethics Code professional supervision, Tarasoff duty-to-warn, and state telehealth practice standard dimensions of suppressed patient distress signal detection.

4. Clinical documentation photograph injection (SilverCloud AI, Quartet Health AI)

Clinical documentation photograph AI processes patient clinical documentation photographs, intake assessment form images, and treatment progress documentation photographs from SilverCloud Health AI (Amwell) at digital therapeutic and integrated behavioural health programme deployments, Quartet Health AI at integrated behavioural health and primary care coordination programme deployments, and integrated digital behavioural health platform clinical documentation management AI tools, extracting clinical documentation accuracy and completeness classifications — clinical assessment completion indicator verifications, treatment goal progress documentation accuracy flags, medication adherence documentation completeness markers, and HIPAA-compliant clinical record integrity verifications — from patient clinical documentation photograph and form image inputs in AI-assisted clinical documentation management pipelines, generating clinical record accuracy alerts, documentation completeness flags, and HIPAA-compliant clinical information management records that behavioural health programme administrators and compliance teams depend upon for HIPAA-compliant clinical record management and integrated care coordination quality assurance. SilverCloud Health AI’s clinical documentation AI processes patient intake assessment and treatment progress documentation photographs through AI-assisted clinical record completeness verification and documentation accuracy tools at integrated employer benefit and health plan digital therapeutic programme deployments where HIPAA-compliant clinical record management is required for covered entity and business associate agreement compliance. Quartet Health AI deploys integrated behavioural health and primary care coordination AI at health plan and provider organisation deployments, processing patient clinical documentation and integrated care coordination records through AI-assisted care gap identification and referral management tools with HIPAA, MHPAEA, and integrated care quality standard dimensions.

The adversarial injection surface is the patient clinical documentation photograph and form image submission pathway: SilverCloud AI or Quartet Health AI patient clinical intake, assessment, and treatment progress documentation photographs submitted through AI-assisted clinical record completeness verification, documentation accuracy assessment, and HIPAA-compliant clinical record management tools for AI clinical documentation integrity classification and integrated care coordination quality determination. An adversarially crafted SilverCloud AI clinical documentation photograph — in which pixel perturbations applied to the clinical assessment completion indicator display, the medication adherence documentation record, or the treatment progress milestone marker in a patient clinical documentation form photograph cause the AI to classify an incomplete or inaccurate clinical record as a complete and accurate HIPAA-compliant clinical documentation record when the actual photograph documents clinical record gaps requiring clinical staff follow-up and documentation completion — can suppress a documentation completeness flag that would otherwise generate a clinical staff follow-up notification and a HIPAA clinical record accuracy audit entry. In integrated digital behavioural health programme environments where SilverCloud AI or Quartet Health AI processes patient clinical documentation photographs across large employer benefit and health plan programme populations without individual clinician review of every AI documentation accuracy classification, adversarial suppression of clinical record completeness flags and documentation accuracy alerts allows incomplete or inaccurate clinical records to be accepted as HIPAA-compliant clinical documentation with integrated care coordination and MHPAEA parity audit consequences.

The HIPAA and MHPAEA consequences of adversarially suppressed clinical documentation accuracy classification in clinical documentation photograph AI span HIPAA PHI accuracy, MHPAEA parity audit, integrated care coordination quality, and state behavioural health programme licensing dimensions. HIPAA 45 CFR §164.310(d)(2)(iii) requires covered entities to implement procedures to ensure the accuracy of PHI stored in electronic information systems; adversarially manipulated SilverCloud AI or Quartet Health AI clinical documentation photograph AI that accepts incomplete or inaccurate clinical records as complete creates HIPAA §164.310(d)(2)(iii) PHI accuracy obligation exposure when adversarially accepted documentation inaccuracies affect patient treatment continuity in integrated care coordination programmes. MHPAEA 29 USC §1185a requires group health plans providing mental health or substance use disorder benefits to ensure that the financial requirements and treatment limitations applied to those benefits are no more restrictive than those applied to medical and surgical benefits; MHPAEA parity compliance audits by DOL and state insurance regulators examine clinical documentation of medical necessity determinations and treatment limitation application — adversarially corrupted Quartet Health AI clinical documentation that suppresses clinical record completeness flags creates MHPAEA parity audit record integrity exposure when documentation gaps affect the evidence base for MHPAEA parity compliance demonstration. State behavioural health programme licensing requirements specify clinical documentation standards for licensed behavioural health programmes; adversarially manipulated clinical documentation AI that accepts incomplete clinical records as compliant creates state behavioural health programme licensing inspection findings with programme licence renewal consequences. Threshold: 50 for clinical documentation photograph AI — reflecting the HIPAA PHI accuracy, MHPAEA parity audit, integrated care coordination quality, and state behavioural health programme licensing dimensions of suppressed clinical documentation accuracy classification.

Integration: mental health and digital health AI image ingestion with Glyphward pre-scan

Mental health and digital health AI image ingestion flows from Spring Health and Lyra Health patient mental health assessment photograph APIs, Crisis Text Line and Woebot crisis support conversation interface screenshot channels, Talkspace and BetterHelp teletherapy session video frame interfaces, and SilverCloud and Quartet Health clinical documentation photograph platforms into PHQ-9 severity and mental health assessment AI, suicidal ideation risk and crisis triage AI, teletherapy session quality and patient distress signal AI, and clinical documentation accuracy and HIPAA compliance AI pipelines. Insert Glyphward’s pre-scan at the ingestion boundary before AI-generated output is committed to mental health severity escalation triggers, crisis counsellor intervention assignments, teletherapy session clinical alerts, or clinical documentation accuracy records:

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

import httpx

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

# Mental health & digital health AI — HIPAA 45 CFR Part 164; APA Ethics
# Code 4.05(b) duty-to-warn; Tarasoff v. Regents common law duty-to-warn;
# MHPAEA 29 USC §1185a; Joint Commission NPSG.15.01.01; 34 USC §290bb-36b
# 988 Lifeline. Suppression of severity classification and crisis escalation
# creates duty-to-warn liability and patient life safety consequences.
THRESHOLD_ASSESSMENT_AI       = 50  # Spring Health/Lyra; PHQ-9; duty-to-warn
THRESHOLD_CRISIS_AI           = 50  # Crisis Text Line/Woebot; NPSG.15.01.01; 988
THRESHOLD_TELETHERAPY_AI      = 50  # Talkspace/BetterHelp; APA supervision; HIPAA
THRESHOLD_CLINICAL_DOCS_AI    = 50  # SilverCloud/Quartet; HIPAA §164.310; MHPAEA


class MentalHealthAIContext(str, Enum):
    ASSESSMENT_AI     = "assessment_ai"      # Spring Health, Lyra — PHQ-9 severity
    CRISIS_AI         = "crisis_ai"          # Crisis Text Line, Woebot — ideation risk
    TELETHERAPY_AI    = "teletherapy_ai"     # Talkspace, BetterHelp — session quality
    CLINICAL_DOCS_AI  = "clinical_docs_ai"  # SilverCloud, Quartet — record accuracy


def threshold_for(context: MentalHealthAIContext) -> int:
    return 50  # all surfaces: patient life safety


async def scan_mental_health_ai_image(
    image_path: str | Path,
    context: MentalHealthAIContext,
    programme_id_hash: str,     # SHA-256 of employer benefit programme or org ID
    patient_session_ref: str,   # e.g. "SPH-SESS-2026-88841", "CTL-CONV-44928"
    assessment_id: str,         # assessment or clinical session identifier
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan a mental health or digital health AI image for adversarial injection
    payloads before forwarding to PHQ-9 severity assessment, crisis suicidal
    ideation risk classification, teletherapy session quality monitoring, or
    clinical documentation accuracy AI systems.

    Raises AdversarialMentalHealthAIImageError if score meets threshold:
      - ASSESSMENT_AI:    threshold 50; HIPAA §164.530(c); Tarasoff; MHPAEA
      - CRISIS_AI:        threshold 50; NPSG.15.01.01; 988 Lifeline; §290bb-36b
      - TELETHERAPY_AI:   threshold 50; HIPAA §164.526; APA Ethics 2.05; state
      - CLINICAL_DOCS_AI: threshold 50; HIPAA §164.310; MHPAEA parity audit
    """
    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": {
                "mental_health_context":  context.value,
                "programme_id_hash":      programme_id_hash,
                "patient_session_ref":    patient_session_ref,
                "assessment_id":          assessment_id,
                "client_scan_id":         client_scan_id,
                "image_sha256":           image_sha256,
            },
        },
        timeout=8.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "programme_id_hash":    programme_id_hash,
        "patient_session_ref":  patient_session_ref,
        "assessment_id":        assessment_id,
        "mental_health_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_mental_health_audit_record(audit_record)

    if result["score"] >= threshold:
        raise AdversarialMentalHealthAIImageError(
            f"Mental health AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"programme={programme_id_hash} session={patient_session_ref}"
        )
    return result


async def write_mental_health_audit_record(record: dict) -> None:
    """Persist audit record to HIPAA-compliant clinical safety documentation store (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialMentalHealthAIImageError(Exception):
    """Raised when a mental health or digital health AI image exceeds the adversarial injection threshold."""
    pass

Call scan_mental_health_ai_image() with MentalHealthAIContext.ASSESSMENT_AI before forwarding Spring Health AI or Lyra Health AI patient mental state assessment photographs to PHQ-9 severity classification and clinical escalation determination AI — the integration point where adversarial suppression of severe depression severity creates Tarasoff duty-to-warn liability and MHPAEA clinical necessity parity compliance exposure, with patient_session_ref linking the Glyphward scan to the patient session record for HIPAA-compliant clinical audit trail documentation. Call with MentalHealthAIContext.CRISIS_AI for Crisis Text Line AI or Woebot crisis support conversation interface screenshots before AI suicidal ideation risk classification and 988 Lifeline emergency response protocol trigger determination, with assessment_id as the crisis triage session identifier for Joint Commission NPSG.15.01.01 suicide risk assessment documentation. Call with MentalHealthAIContext.TELETHERAPY_AI for Talkspace or BetterHelp teletherapy session video interface frames before AI patient distress signal detection and therapist supervisor escalation, with patient_session_ref for HIPAA-compliant teletherapy session clinical record and APA Ethics Code professional supervision documentation. Call with MentalHealthAIContext.CLINICAL_DOCS_AI for SilverCloud or Quartet Health clinical documentation photographs before AI clinical record completeness and HIPAA documentation accuracy assessment, with programme_id_hash for MHPAEA parity audit and HIPAA §164.310 PHI accuracy compliance documentation. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four mental health and digital health AI injection surfaces; covers how adversarial pixel-level perturbations cause AI misclassification without detectable visual artefacts at human review resolution.
• Healthcare AI prompt injection — related regulatory framework covering HIPAA, patient safety, and clinical AI integrity applicable to mental health and digital health platform contexts.
• Telehealth and remote patient monitoring AI prompt injection — adjacent clinical AI injection context covering HIPAA, patient safety monitoring, and RPM platform adversarial image attack surfaces.
• Free tier — 10 scans/day, no card required — start scanning mental health and digital health AI patient images at development volumes before committing to a production plan.