Citt Safety Standard

Safety is the product,
not an afterthought.

Mental health AI carries real stakes. Citt.ai was built with crisis detection, human oversight, and clinical accountability at its foundation, not bolted on later.

Operating targets

95%
Crisis detection sensitivity
operating target
≤5%
False positive rate
operating target
100%
Jailbreak resistance
evaluation target
0.92
F1 score
operating target

Targets are evaluated via automated adversarial test suite (CEP v2). Automated adversarial test suite (CEP v2)

Latest measured performance

These figures come from the latest persisted CEP v2 evaluation run, not marketing copy.

No persisted CEP run is available yet. The runtime still uses the operating targets above, and the safety evaluation cron will publish measured results after it runs.

The Citt Safety Architecture

Four layers that work together to ensure no patient in distress falls through the cracks.

Crisis Detection on Every Message

quickRiskCheck() runs before AI processing on web chat, stream, secure-route, media captions, feedback, WhatsApp, and multi-agent patient paths. It covers suicidal ideation, self-harm, abuse, and explicit crisis language.

web chat, stream, secure-route, media captions, feedback, WhatsApp, multi-agent

Human in the Loop

Every AI output carries provenance and confidence scores. Therapists review, approve, or override AI-generated clinical content. Crisis events trigger immediate therapist notification.

Therapist approval workflow on all AI clinical outputs

Full Audit Trail

Every safety decision, crisis event, and clinical action is logged with timestamps, user IDs, and context. Immutable audit log for clinical accountability and regulatory review.

Audit logging active on all clinical actions

Privacy-Sensitive Data Handling

Encryption, access controls, audit logs, and vendor controls support privacy-sensitive care environments and HIPAA-regulated deployments.

Access restrictions and audit logging in place
Adversarial evaluation

Tested against attacks designed to defeat it

Our CEP v2 evaluation framework includes dedicated test cases for the failure modes that have caused harm elsewhere: obfuscated crisis language, jailbreak attempts, and gradual escalation patterns. Obfuscation and gradual escalation are evaluated here; the live hot path remains deterministic.

Obfuscation resistance
Evaluation target for disguised language
90%
Jailbreak resistance
Resists prompt injection attacks
100%
Gradual escalation detection
Evaluation target for slow-build patterns
85%
Known limitations

What the live runtime does and does not claim

The hot-path safety gate is deterministic keyword and pattern detection. It does not depend on an LLM to decide whether to block a crisis message.
Obfuscation and gradual escalation are evaluated in CEP v2, but they are not claimed as fully solved runtime capabilities until deterministic support and measured results are present.
Crisis events create therapist alerts and audit rows. Elevated risk signals are logged separately as non-alertable timeline events.
Feedback, WhatsApp, multi-agent, media, stream, and secure-route patient paths now use the same crisis escalation semantics before normal AI processing.

How patient records are created

Therapists can run their full practice in Citt.ai, including for patients who have not yet signed in to the patient portal. We gate every patient-facing feature behind an explicit claim step and keep AI access linked to a therapist or clinic that is responsible for the clinical purpose and oversight model.

Managed records

A therapist on Plus or Full Access can add a patient by name and email. The record supports scheduling, notes, billing, and transcription. Until the patient claims, no chat, check-ins, assessments, or WhatsApp messages are sent to them.

Claim attaches identity

When the patient is ready, the therapist sends a claim invite. The invite is a 14-day, single-use, HMAC-hashed token. Claiming attaches authentication credentials to the existing record. We do not create a second account or duplicate the data.

Cross-tenant safety

If the patient's email is already on another therapist's roster, the new therapist gets the same response shape as a brand-new creation. No cross-tenant data is revealed. Every cross-tenant link is logged at high severity, the patient is notified where they have claimed, and prior therapists receive a heads-up.

Therapist-led access and joint controllership

For managed records and patient-facing AI flows, the therapist or clinic determines the clinical purpose while Citt.ai determines essential platform means such as hosting, AI infrastructure, safety tooling, subprocessors, and retention architecture. The patient-facing chat, check-ins, and assessments are not standalone Citt.ai care; they stay linked to a clinician or clinic that keeps the patient active for regular therapy, maintenance support, or a similar oversight arrangement.

Data Practices

We are transparent about who processes your data and why. We maintain contractual and privacy terms for the services listed below and provide DPA materials for customer review on request.

Sub-processorPurposeData CategoriesLocationSafeguard
AWS (database and storage)Database and object storageAccount data, patient-care data, files, logsUS / EUEncryption, access controls, contractual terms
OpenAIAI chat responses, transcriptionConversation content, prompts, transcription payloadsUSContractual controls, no training on API customer data by default
AWS (application hosting)Application hostingInfrastructure processing and operational logsUS / EUEncryption, access controls, contractual terms
StripePayment processingPayment tokens, subscription metadataUSPCI DSS Level 1
GoogleDemo scheduling, optional calendar integrations, and consent-gated website measurementDemo-booking contact details, calendar metadata, device identifiers, and consented website measurement dataUS / EUConsent gating where applicable, contractual terms, and vendor transfer commitments
Meta (WhatsApp Business)Messaging channelMessages (when opted in)US / EUPlatform terms, DPA commitments
ResendTransactional emailEmail addresses, notification contentUSSOC 2, TLS encryption
MailgunGTM outbound emailEmail addresses, marketing contentEUSOC 2, GDPR DPA
DeepgramReal-time transcriptionAudio streams where transcription is enabledUSContractual restrictions and security controls

Evaluating Citt.ai for your organisation?

We provide full technical documentation, safety architecture whitepapers, and can arrange a clinical review for health system procurement teams.