Security
Where your data goes, who processes it, and which controls and terms govern it
Atteston processes two sensitive things: your security evidence and your customers' questionnaires. This page states where that data goes, which providers process it, and which controls and contractual boundaries govern it. Deployment-specific terms are available through a security review.
Data-handling boundaries
Atteston does not train models on customer content
Atteston does not train its own models or assemble training datasets from your evidence and questionnaires. Configured inference providers process the task inputs described below; their training, retention, and data-use restrictions depend on the deployment configuration and contractual terms.
Tenant boundaries are enforced in the application and database
Membership checks, explicit account context, transaction-local database context, and row-level security are layered to prevent cross-account access. Automated checks exercise the runtime role and cross-tenant isolation paths.
Workbook writes are approval-bound
The approved-answer write and export path selects account-scoped approved answer versions. Questions that still need review or input remain in the review workspace instead of being treated as approved output.
Weak or missing support stays visible as review work
Questions with no admissible evidence are grouped Needs you and shown as not found in the sources searched. They are not elevated into the gated verified-reuse path.
Controls
| Control area | Implementation |
|---|---|
| Tenant isolation | PostgreSQL row-level security scoped by account, enforced in the database and verified by an automated guard at service startup. |
| Encryption in transit | Public client-to-service traffic in the hosted service uses TLS. Hosted database and Redis connections also use TLS. Some private application service-to-service calls use HTTP endpoints restricted to the VPC; deployment-specific transport boundaries are available through a security review. |
| Encryption at rest | In the hosted service, primary database volumes and object storage are encrypted at rest with provider-managed encryption. Customer deployments may use customer-managed KMS for object storage. |
| Audit trail | The append-only audit ledger defines event types for selected job outcomes, denials, privileged operations, and sensitive actions. Answer revisions are preserved separately as immutable answer versions; verification decisions and reviewer actions use their own durable records. |
| Access control | Role-based access within workspaces; membership is account-scoped and managed by your team. |
| Data deletion | Account deletion and member-erasure workflows are supported. Document-level self-service deletion is not currently available. |
Subprocessors
Default providers form the hosted launch baseline. Conditional providers receive data only when the named feature is used. Optional providers are inactive unless the deployment provisions and configures them. Each row states the applicable task category and data shape. Adding another provider requires updating this disclosure and customer notice under the data processing addendum before customer data flows to it.
| Provider | Availability | Purpose | Data it receives |
|---|---|---|---|
| AWS | Default | Hosted platform (compute, database, storage, logs); Textract (OCR, disabled by default until launch enablement) | Customer content and operational metadata, encrypted at rest; transport follows the encryption-in-transit boundaries in the controls table above. |
| Anthropic PBC | Default | Answer drafting (primary and draft fallback models), answer verification, evidence claim extraction, knowledge summarization, and question interpretation when enabled, over the first-party API | Question text, prompts, draft answers, selected evidence and document text, the full text of ingested evidence chunks (for claim extraction), document and table metadata, sample rows used for summaries, and token usage metadata. |
| AWS Bedrock | Default | Embedding generation | Full text of ingested document chunks; model-generated document summaries, table summaries and extracted claims derived from them; and customer retrieval queries. |
| Stripe | Default | Billing | Billing identifiers and subscription state only. No questionnaire or document content. |
| AWS SES | Default | Transactional email | Recipient address and message content for sign-in, invitation, and account email. |
| PostHog | Optional | Product analytics | Product usage and event metadata, within a telemetry redaction policy. No raw customer source text. |
| OpenAI | Optional | LLM inference for configured evaluation, reranking, and OCR post-processing lanes | Prompt inputs required by the configured task. Embedding text is limited to explicit evaluation or comparison use, not automatic production embedding fallback. |
How the AI is constrained
Evidence-linked drafting
Generated answers can carry citations to selected passages from your uploaded evidence. Weak or missing support routes the item to evidence review or manual input instead of the verified-reuse path.
Five explicit review groups
Every question sits in one of five groups: Needs you, Ready to review, To confirm, Approved, or In progress. Verified reuse is reserved for gated exact reuse. A draft from your documents with admissible citations to current evidence is grouped Ready to review or To confirm and waits for a person's approval; a question with no admissible evidence, an imported answer, or input the owner must give is grouped Needs you.
Approval-bound output
The review workspace lets people accept, edit, reject, or route work for more input. Workbook writes and exports use approved answer versions.
Durable, separated records
Answer text and revisions live in immutable answer versions. Verification decisions and reviewer actions have dedicated records. Selected security-sensitive and operational events use the append-only audit ledger.
Compliance posture, honestly
Atteston follows SOC 2-aligned control practices — tenant isolation, least-privilege access, and audit logging — but does not currently have a SOC 2 report. We will not claim an attestation we have not completed, and this page will state it plainly when that changes. A data processing addendum and security review materials are available through the contact route.