Security.
How we protect your data.

Uploaded documents, exports and backups are stored in AWS ap-south-1 (Mumbai, India) with server-side encryption at rest (AES-256). All traffic is encrypted in transit with TLS 1.2 or later. LexVio staff do not read your documents — access is restricted to logged security or abuse investigations, or support you ask for, per our DPA. Audit logs are append-only, enforced by database triggers, and kept for at least seven years. India holds no EU or UK adequacy decision; the transfer terms are in the DPA.

Every model vendor is reached through a gated client. An air-gapped deployment disables every external model vendor and all telemetry with one setting. User-supplied text is screened for prompt injection before any model sees it, on every vendor path. Each request carries input and output token caps, and each organisation has a daily AI-spend ceiling that stops further model calls when it is reached. Repeated prompt context is cached so the same document is not re-sent in full on every turn. Each Anthropic request is routed by task across Claude Haiku 4.5, Sonnet 5, Opus 5 and Fable 5.1; an organisation on zero-data-retention terms is served from Opus rather than Fable, because Fable requires 30-day retention. Under Anthropic's, OpenAI's and Google's API terms, your content is not used to train their models; we do not train models on your data without explicit consent.

Your administrators set the rules for your organisation. Access policies — an IP allowlist, the platforms allowed (web, phone), and a maximum session age counted from sign-in — are checked on every authenticated request except sign-in and sign-out, and a change that would lock out the person saving it is refused. Multi-factor authentication is not enforced today, and we do not claim it. AI administration lets an administrator lower the model tier the platform would otherwise use, switch individual AI features off, and set a daily AI-spend cap; prompt-injection hits and blocked requests are kept as security events with a content hash, never the prompt, and the AI request history records the request type, user, model and time, not the content. Every change to a matter, a document, a workflow step or an admin setting writes an append-only audit row. Documents classified as identity documents are never sent to an AI vendor and never indexed. Exported minutes, Vio answers and redlined documents carry machine-readable AI-provenance metadata in the file; nothing is printed on the page. Credentials for your own email relay, SMS or WhatsApp account are stored encrypted and never shown back.

Aligned with applicable data protection regulations; independent assessment not yet completed. SOC 2: controls mapped and evidenced in-product, not yet audited. ISO 27001: not yet certified. Automated dependency and static-analysis scans run weekly; a third-party penetration test has not yet been performed. Current status on /trust; the full measures and sub-processor list are in the DPA.