Security

All accounts are protected by industry-standard authentication with encrypted JWT session tokens. Passwords are hashed using bcrypt with per-user salts — we never store or transmit plaintext credentials.

Brute-force protection is built in: after 5 failed login attempts, the account enters a timed lockout period. Email verification with expiring codes prevents account impersonation. OAuth login via Google is also supported.

Admin endpoints are gated by a separate authorization layer — standard user tokens cannot access administrative functions regardless of how the request is crafted.

All traffic is encrypted in transit via TLS 1.2+ with HSTS preloading enforced across the entire domain. Sensitive cookies are encrypted at rest using AES-256-GCM with unique initialization vectors per value.

Database backups use point-in-time recovery with encrypted snapshots retained for disaster recovery.

Every request is authenticated before it reaches any backend service, and per-user message limits are enforced server-side — a modified client cannot raise its own allowance.

Secrets and API keys live in encrypted environment variables, scoped to the services that need them. Application code runs on managed infrastructure with no long-lived servers for an attacker to persist on.

Conversations are stored in a private database with row-level security enforced at the database layer. Each user can only access their own data — this is enforced by the database engine itself, not just application logic.

We do not sell your conversations, and we do not use them to train models. Your messages are sent to the AI providers named in our Privacy Policy for the sole purpose of generating a reply, under API terms that exclude training on that content. No analytics pipeline processes your messages, and no data broker has access to them.

Secrets and API keys are managed through encrypted environment variables — never hardcoded, never committed to source control, never logged.

The application enforces a strict Content Security Policy that prevents cross-site scripting (XSS), clickjacking, and unauthorized resource loading. Frames are blocked entirely. Form actions are restricted to the origin domain.

Additional hardening headers are set on every response: X-Content-Type-Options, X-Frame-Options, Referrer-Policy, and a Permissions-Policy that disables camera, microphone, and geolocation access by default.

All user input is validated against strict schemas before processing. Message content, file uploads, and API parameters are type-checked and length-bounded using schema-based validation at every entry point.

File uploads undergo dual validation: MIME type whitelist and magic byte verification. This prevents content-type spoofing — a file claiming to be an image but containing executable code will be rejected at the byte level.

All payments are processed through Stripe — a PCI DSS Level 1 certified payment processor. We never see, store, or transmit your card number. The full payment flow happens on Stripe's infrastructure.

Webhook events from Stripe are verified using cryptographic signature validation before any change to your subscription is applied. Entitlements are derived from the verified Stripe event — not from client-submitted data — so access cannot be granted by a forged request.

While the platform is uncensored, it is not unmoderated. A multi-layer content safety system operates in real time:

• Pre-flight moderation screens inbound messages against established safety classifiers
• Output filtering detects and removes repetitive or degraded model responses in-stream
• Post-generation review scores outputs for safety compliance and applies surgical corrections when needed
• Transcript anomaly detection catches model drift patterns (dialogue simulation, affirmation loops) before they reach the user

Eight non-negotiable safety boundaries are enforced at the system level — see our FAQ for the full list.

Production logs are watched continuously by an automated rule set that pages on error patterns — failed logins, provider outages, and unexpected exceptions — rather than waiting for a user to report them.

Alarms trigger on provider health check failures, billing anomalies, and unexpected error rates. The system is designed to fail safe: if a component becomes unreachable, the default behavior is to block the action rather than allow it through unchecked.

The real-time research tools (academic search, court records) query live external databases and return only verified sources. Citations are never fabricated by the model — they are fetched from authoritative indexes and presented with verifiable references.

Tool invocations are deterministic: the AI cannot choose to hallucinate a source when a real tool is available. When a query matches a research or legal pattern, the system mandates tool use before generating a response.