Skip to content

Trust center

Your work,under your control.

How LiteSurface protects projects, keys, and decisions. We describe what the product does today and mark anything still in progress as planned.

01

Data ownershipProject-centric by design.

Your research, concepts, and decisions belong to you, organized so they can move as a unit. Nothing about how LiteSurface stores data ties your work to a single person’s account.

  • Projects are the unit of ownership

    In product

    Concepts, sources, claims, evaluations, experiments, decisions, and artifacts all belong to a project. Transfer a project and the whole dataset moves with it.

  • Tenant scoping on every record

    In product

    Every tenant-owned row carries its workspace. Services resolve each entity to its workspace and check membership; an id in a URL is never enough.

  • Cross-tenant tests

    In product

    Automated tests seed two workspaces and attempt to read each other’s data through the API, so isolation is checked on every change.

  • Export and deletion

    In product

    Projects can be exported. Deleted projects are purged, along with their stored files, 30 days after deletion.

02

Keys and providersYou decide who sees what.

LiteSurface does not resell model access. It calls providers with your keys, under your policy, and every provider involved is named on this page. See Governance for the admin controls.

  • Bring your own keys

    In product

    Each workspace can connect its own OpenAI and Anthropic keys. Keys stay server-side: they are never returned to the browser and never placed in model context.

  • Allowed providers per workspace

    In product

    Admins choose which providers a workspace may use. Routing and failover respect that list, so content never reaches a provider you have excluded.

  • Only what the task needs

    In product

    Prompts are assembled from the specific fields a task requires rather than whole projects. Application code refers to model aliases, never to vendors directly.

  • Log redaction

    In product

    Logs automatically redact authorization headers, cookies, keys, secrets, tokens, prompts, raw model responses, and signed URLs.

  • Sample project and fixture mode

    In product

    Explore a sample project before you run your own. Self-hosted deployments can also run in fixture mode, which replaces every model call with deterministic outputs.

  • Field-level sensitivity policy

    In product

    Fields classified as sensitive are never sent to model providers: they are removed before any prompt is assembled.

03

What goes whereEvery recipient, named.

The services that receive data when the product runs, what they receive, and when. The subprocessor list covers the companies behind them.

Recipients of data, what they receive, and when
RecipientWhat it receivesWhenNotes
OpenAITask prompts built from the fields a task needs, such as strategy, concept genomes, cited claims, and source text for extraction. Claim text for embeddings.Tasks routed to OpenAI, including one of the two evaluations and embeddings for retrievalRouting is configuration; failover follows your allowed-provider list.
AnthropicTask prompts of the same kinds, built from the fields a task needs.Tasks routed to Anthropic, including the other of the two evaluationsEvaluates independently; never sees the OpenAI evaluator’s answer.
Search providerSearch queries derived from the research focus. Not documents, claims, or concepts.Research and evidence-gap runsTavily or Brave, as configured.
Public websitesOrdinary page requests for the sources being fetched.When you add a URL, or a search result passes relevance triageFetched through the SSRF-safe fetcher; robots.txt is honored.
Email deliveryRecipient address and message content.Sign-in, invitations, and notificationsResend in production; a local mail catcher in development.
Object storageExports, artifacts, and stored files.Artifact export, project export, attachmentsTenant-prefixed paths; downloads use short-lived signed URLs.

Nothing is sent before you allow it

Project content reaches OpenAI or Anthropic only after a workspace admin accepts the provider disclosure and allows that provider. Self-hosted deployments can run in fixture mode, where no prompt leaves your servers.

04

Encryption and application securityDefaults, not add-ons.

  • HTTPS everywhere

    In product

    Production traffic is served over TLS with HTTP Strict Transport Security, and insecure requests are upgraded.

  • Security headers

    In product

    The site and API send a Content Security Policy that denies framing, along with nosniff and a strict referrer policy. The site also sends a restrictive permissions policy.

  • Secure sessions

    In product

    Authentication uses secure, HTTP-only session cookies. Cross-origin requests are limited to the product’s own origins.

  • Encryption at rest, documented

    Planned

    A published description of at-rest encryption for the hosted service, including databases, object storage, and backups.

05

Access rolesLeast privilege, four levels.

Workspaces have four roles. Projects add their own owner, editor, and viewer roles for sharing individual projects. Permissions are checked in the service layer on every request.

Capabilities of each workspace role
CapabilityOwnerAdminMemberViewer
Read workspace contentAllowedAllowedAllowedAllowed
Create and edit projectsAllowedAllowedAllowedNot allowed
Run AI workflowsAllowedAllowedAllowedNot allowed
Record decisions and outcomesAllowedAllowedAllowedNot allowed
View usageAllowedAllowedAllowedNot allowed
Manage membersAllowedAllowedNot allowedNot allowed
Manage prompts and providersAllowedAllowedNot allowedNot allowed
Delete the workspaceAllowedNot allowedNot allowedNot allowed

06

Audit logWho did what, and when.

Consequential actions are written to an audit log with the actor and time, and admins review it in the separate admin console.

Recorded events include

  • Membership changes, invitations, and role updates
  • Project creation, duplication, transfer, and deletion
  • Strategy and scorecard changes
  • Portfolio decisions and experiment outcomes
  • Workspace AI policy and budget changes
  • Provider breaker resets and prompt activations
  • Exports created and downloaded

07

Budgets and provider breakersSpend and failure, contained.

  • Budgets that stop runs

    Each workspace has a spending budget. When it is exhausted, new runs are refused instead of overspending, and the usage ledger shows cost per project and run.

  • Provider circuit breakers

    Repeated failures take a provider out of rotation. After a cooldown, a single probe call tests it before traffic returns. Admins can reset a breaker, and resets are audited.

  • Visible degradation

    If one evaluator is unavailable, evaluation continues with the other and the result is marked partial, with lower confidence.

  • Prompt versions with rollback

    Prompts are versioned assets. Admins activate or roll back a version, and every result records the prompt version that produced it.

08

Prompt injection defensesRetrieved pages are treated as data.

A research tool reads pages written by strangers, so it must assume some of them contain instructions aimed at the model. These defenses apply to every run, and a prompt-injection evaluation suite is part of the release checks.

  • A fixed trust order

    System policy, then operator prompts, then the signed-in user, then project state, then retrieved content, then model output. Lower levels can never redefine higher ones.

  • Retrieved pages are data

    Fetched content is wrapped and labeled as untrusted in every prompt. Instructions inside it are treated as text to analyze, never as commands.

  • Tool allowlists

    Each AI task gets a fixed set of tools. Claim extraction has none, and no model is given arbitrary network access.

  • No secrets in context

    Keys and credentials are never placed in model context, so there is nothing for an injected instruction to exfiltrate.

  • Validated citations

    Model outputs are checked against schemas, and citations to claim ids that do not exist in the project are dropped.

  • SSRF-safe fetching

    The fetcher blocks private, loopback, link-local, and cloud metadata addresses, re-validates every redirect, pins the resolved address, and caps size and time.

09

On the roadmapPlanned, not promised as shipped.

These items are not available today. We will record them in the changelog when they ship.

  • SSO and SCIM provisioning for Enterprise workspacesPlanned
  • Published at-rest encryption details for the hosted servicePlanned
  • Data residency options for Enterprise deploymentsPlanned

10

Responsible disclosureTell us, and we will work with you.

If you believe you have found a security vulnerability, email security@litesurface.com with a description, steps to reproduce, and the impact you observed. We will acknowledge your report, keep you updated, and credit you if you wish.

While researching, please

  • Only access data in accounts you own or have permission to use
  • Avoid degrading the service, including denial-of-service testing
  • Do not use social engineering or physical attacks
  • Give us reasonable time to fix an issue before disclosing it

Security contact

For vulnerability reports and security questionnaires. Enterprise customers can request our security documentation and DPA.

security@litesurface.com

11

PoliciesThe legal detail.

Review it first,then allow a provider.

Explore the sample project, read the provider disclosure, and allow OpenAI or Anthropic only when your review is done.

  • Solo and Team are free during early access
  • Explore a sample project first
  • Bring your own OpenAI and Anthropic keys