Resources
Security is a launch requirement, not a roadmap item.
Rounic processes some of the most sensitive information a private equity firm holds — live deal processes, valuation assumptions, management information, investment decisions, portfolio results, and potentially material nonpublic information. The controls below are product requirements, not upsells.
Permissions inherited, never broadened
- Access decisions are made before retrieval, against tenant, user, group, source, object, and classification
- A user is never shown a snippet from a document they could not open at the source
- “No result” is phrased so it cannot confirm the existence of restricted material
- Anything derived from a restricted source inherits that restriction
Tenant isolation
- Data, caches, vector indexes, and search indexes are isolated per tenant
- One firm’s corpus can never be retrieved into another firm’s answer
- Single-tenant deployment available where security requirements justify it
Encryption and secrets
- Encrypted in transit and at rest
- Secrets held in a managed secret store, never in configuration or logs
- Credentials redacted from telemetry and error paths
- Customer-managed keys available for enterprise deployments
Model-provider controls
- Approved providers only, under contracts that prohibit training on your data
- Classification-aware egress — the most sensitive material can be pinned to stricter processing paths
- Private-model and private-networking options for firms with strict MNPI policies
- Your data is never used to train models shared with anyone else
Audit
- Every query, every source touched, and every answer recorded
- Administrative changes recorded separately and immutably
- Exportable audit trail for compliance and recordkeeping review
- Legal hold and retention policy configurable per tenant
Agent and injection safety
- Text inside an ingested document is data, never an instruction to the system reading it
- Tools are capability-scoped with validated arguments and an egress allowlist
- Writes to external systems require human approval
- Uploads scanned for malware before processing
Assurance programme
What we run, and where we are.
- Secure SDLC with dependency, container, and secrets scanning
- SAST/DAST and infrastructure-as-code review
- Threat modelling across ingestion, retrieval, the model gateway, tool execution, and admin paths
- Annual penetration testing at minimum, and after major architecture changes
- Incident-response runbook with tabletop exercises
- Vendor and subprocessor inventory, access review, and backup/restore testing
Rounic is early. We build against SOC 2 controls from the start, but we will not claim a certification we do not yet hold, and we will not promise a data-residency region until every subprocessor path on it has been verified. If your diligence process needs our current status in writing, ask — we will send you the real answer.
Security is part of the UX
A control the user cannot see is a control the user cannot trust. Rounic makes the boundary visible: which sources an answer used, which it could not reach, why a document is unavailable, and who has seen what. Permission-denied states explain the shape of the restriction without leaking its contents.