Security and Trust Center

How Custard protects your research data

This page is Custard's Trust Center: the controls in place today, stated so a reviewer can verify them claim by claim - the data the platform is and is not built to hold, the mechanisms that enforce each control and how they fail, the providers that process customer data, and the review materials available on request. It is written for university IT, security, and procurement reviewers conducting a vendor assessment, and for the research team leads who answer to them.

Use Custard only for approved data

Custard is built for pre-commercial research validation. The approved categories below define what the platform is designed and controlled to hold; the not-permitted categories are deliberate scope statements, not gaps we plan to close. Data outside the approved scope belongs in your institution's approved systems.

  • Approved: unpublished research descriptions and invention disclosures, discovery interview recordings and transcripts, stakeholder and prospect business-contact information, project notes and uploaded research documents, and the analyses Custard generates from them.
  • Not permitted: protected health information (PHI). Custard is not designed to hold PHI or any data covered by HIPAA.
  • Not permitted: regulated clinical or human-subject data. Identifiable clinical trial data and regulated human-subject research data belong in your institution's approved systems. Discovery interviews about a problem space are in scope; regulated study data is not.
  • Not permitted: CUI and export-controlled data. Controlled Unclassified Information (NIST 800-171) and export-controlled data (ITAR/EAR) must not be uploaded.
  • Not permitted: government IDs and financial accounts. Social security numbers, bank or financial account numbers, and payment-card numbers must not be uploaded. The DLP scanner blocks AI prompts containing SSN or payment-card patterns outright - the request fails rather than being redacted and forwarded.
  • Not permitted: student education records (FERPA). Custard is not a student-information system. Do not upload student education records beyond the research-validation use this page describes.

The same security for every project

There is no lighter version of Custard's security. Every project, every request, and every background job runs behind the same server-enforced protections - a server-side control, not a UI toggle. Enforcement happens in the backend before any external call is attempted: a disallowed call fails closed with a 403 before a single byte reaches an outside vendor, regardless of what the interface shows. The same rules are enforced at one network choke point, so background jobs and autonomous features are governed exactly like live requests.

  • Tenant isolation verified on every read and write. Project membership is checked on the server for every project-scoped operation, and a static audit test fails the build when a new route that takes an ID lacks an access check.
  • A single outbound gate. Every outbound call passes through a single server-side choke point that scans prompts for secrets and PII, blocks disallowed destinations, and applies the same rules to background jobs as to live requests.
  • An append-only audit trail. Security-relevant events are recorded in an audit trail with no API to edit or delete entries; the scheduled 365-day retention sweep is its only deletion path.
  • Zero-retention AI, never trained on your content. Platform AI runs through Amazon Bedrock under a zero-retention arrangement: prompts and outputs are not stored by the provider and never train any model - ours or anyone else's.

How we protect data in transit and at rest

  • SSRF-hardened outbound traffic. All server-side outbound HTTP goes through a single hardened fetch layer that resolves every destination hostname, rejects private, loopback, link-local, and cloud-metadata addresses, and re-validates every redirect hop.
  • Secrets stripped from logs. Credentials, tokens, session cookies, connection strings, and signed URLs are redacted from all server logs before writing.
  • AES-256-GCM for stored keys. Stored API keys are encrypted with AES-256-GCM. Decryption failures fail closed.
  • Governed uploads. File uploads require an authenticated, email-verified session, pass a MIME-type allowlist and size caps, and every uploaded object is private by default.

How we handle AI

  • Zero-retention platform AI. The platform default model is served through Amazon Bedrock under a zero data retention arrangement. Your prompts and the model's outputs are not used to train any model.
  • Bring your own LLM, safely. If you configure your own AI provider, your API key is encrypted at rest with AES-256-GCM, is never sent to the browser, and is never written to logs. Processing then runs under your agreement with that vendor, not ours.
  • A DLP guard on every prompt. Every AI prompt is scanned for secrets and PII patterns before it leaves the server. A call that trips the scanner is blocked outright - not redacted and sent anyway.
  • Governance follows background work. The same server-side security rules flow through all background AI jobs - not just live requests.
  • No AI runs on page load. All AI generation is explicit: triggered by an action you take or a background job you configured.

AI provider and sub-processor register

These are the third-party providers that may process customer data, the purpose each serves, and the data-use terms that apply. This register mirrors the sub-processor list in the Privacy Policy, which is the authoritative version and where we give at least 30 days' notice of new sub-processors.

  • Amazon Web Services (Amazon Bedrock). Large language model inference for platform AI features (United States). Zero-retention endpoint: prompts and outputs are not retained, saved, or sold, and never train any model.
  • Replit, Inc. Cloud hosting, application runtime, PostgreSQL database, and object storage (United States).
  • Google LLC. Calendar, Meet, and Sign in with Google - only when you connect Google (United States, EU).
  • Microsoft Corporation. Microsoft Graph for Outlook Calendar, Outlook Mail, and Teams - only when you connect Microsoft (United States, EU).
  • Zoom Video Communications, Inc. Meeting creation - only when you connect Zoom (United States).
  • Postmark (ActiveCampaign LLC). Transactional and outreach email delivery (United States).
  • Exa AI, Hunter.io, Apollo.io, ScholarAI. Prospect discovery and enrichment data providers. Requests to them pass through the same server-side outbound gate as every other external call.
  • Bring-your-own AI provider. If you configure your own OpenAI-compatible or Anthropic endpoint, that processing runs under your institution's agreement with that vendor, not ours.

What we keep, for how long, and how deletion works

  • Deleted content: gone from backups within 30 days. Content you delete in the product leaves active systems immediately and is purged from encrypted backups within 30 days.
  • Account deletion: by request, up to 90 days. Request account deletion at privacy@mira-insight.com. Account data is retained for as long as your account is active, plus up to 90 days for backups and audit logs.
  • Prospect PII: 180 days. Contact records fetched from external research services that were never engaged in outreach are purged 180 days after creation by a scheduled sweep. Active outreach relationships are preserved.
  • Abandoned interviews: 365 days. Interview bookings that ended as declined, cancelled, or no-show are purged after 365 days. Completed and still-scheduled interviews are kept as active research records.
  • Audit trail: 365 days, append-only. The application audit trail is retained for 365 days. There is no API to edit or delete audit events; the scheduled retention sweep is its only deletion path.
  • Project deletion is immediate and cascading. Deleting a project removes its hypotheses, transcripts, evidence, wiki, analyses, interviews, and campaign data in the same operation.
  • Uploaded files are project-bound and swept. Uploaded objects are bound to their project by access policy. When a file reference is removed, the backing object is deleted, and an orphan sweep runs every 6 hours.

For universities and research institutions

University IT, procurement, and research-compliance reviewers evaluate vendors against specific higher-ed frameworks. This section maps the controls above to those checklists - and is equally plain about what Custard is not intended for. Every claim here describes a control that already exists on this page; none is aspirational.

  • Vendor review, organized the way HECVAT-style assessments ask. Data protection: encryption in transit and at rest, zero-retention AI processing. Access control: tenant isolation verified on every read and write, server-side authorization checks tested in our build pipeline. Retention and disposal: fixed retention sweeps, immediate cascading project deletion, and deletion on request. Third parties: a named AI processing path (Amazon Bedrock) and the full sub-processor register above. Reviewers can request the materials listed below, including a completed HECVAT 4.1.6 response.
  • Human-subjects and IRB data. Interview transcripts and recordings stay inside your project and are encrypted at rest. AI processing runs under the zero-retention arrangement above. Participant names in distilled research records are reduced to initials, deleting a project removes its transcripts, evidence, and interviews in the same operation, and fixed retention sweeps purge abandoned interview records after 365 days. These are the controls a researcher can cite in an IRB protocol's data-management plan.
  • FERPA applicability. Custard is a research-validation tool, not a student-information system, and we make no claim of FERPA compliance. Customers should not upload student education records. Where institutional data is involved, the controls that support a school-official-style review are the ones on this page: tenant isolation on every request, no training on your content, and deletion on request.
  • Fit with university AI policies. No training on institutional data - by us or any third party; no vendor retention - platform AI runs through a named processing path, Amazon Bedrock, under a zero-retention arrangement; and a bring-your-own-LLM option so your institution can route AI processing under its own vendor agreement instead of ours.
  • What Custard is not intended for. Custard is not intended for export-controlled data (ITAR/EAR) or Controlled Unclassified Information (NIST 800-171), and does not process GLBA-covered financial-aid records. These are scope statements, not gaps we plan to close - use institution-approved systems for that data.

University security-review materials

For a formal vendor or security review, the following materials are available on request through the support page. We answer with specifics.

  • Completed vendor-assessment answers. Responses organized the way HECVAT-style higher-ed assessments ask, mapped to the controls on this page - including a completed HECVAT 4.1.6 response.
  • Data Processing Addendum (DPA). For customers on plans with a DPA, including direct notice of sub-processor changes to the contract contact.
  • Security architecture summary. A written summary of the controls described on this page - authorization model, encryption, outbound-gate design, and audit trail.
  • Data-flow and sub-processor documentation. What data goes to which provider, for what purpose, and under what terms - matching the register above and the Privacy Policy.
  • Retention and deletion schedule. The schedule on this page in a form a reviewer can attach to an assessment.

Report a security concern

If you believe you have found a security vulnerability in Custard, report it to privacy@mira-insight.com with enough detail to reproduce the issue. Please do not access data that is not yours or disrupt the service while testing. We review every report and respond with what we find. Privacy questions and data-deletion requests go to the same address.

What we don't do

  • We do not train models on your research data. Platform AI runs under a zero-retention contract.
  • We do not store LLM API keys in plaintext. Keys are AES-256-GCM encrypted at rest.
  • We do not let a disallowed call reach an external vendor. Enforcement is server-side at the network choke point.
  • We do not log raw prospect PII or credentials. Email addresses in logs are masked and person names in distilled research records are reduced to initials.
  • We do not expose one tenant's data to another. Every project-scoped read and write verifies membership on the server.
  • We do not auto-run AI on page load. All AI generation is explicit.

Questions about a specific control, reviewing Custard for your team, or preparing a vendor assessment for your institution? Reach us through the support page and we will answer with specifics, including our completed HECVAT 4.1.6 assessment answers for formal reviews.

Last reviewed: August 17, 2026

Privacy Policy · Terms of Use