Legal
Data and Security Terms
The controls that are actually in place, the remaining risks, and the claims Lumevo does not make.
Last updated 12 September 2026. Questions: support@lumevo.digital
These terms describe the controls that are actually in the Lumevo workspace as of September 2026, and the risks that are still open. They are for customer security review. They are not a certification, a pentest report, or a promise that the remaining gaps have been closed.
1. Claims we do not make
- SOC 2 Type II — not completed
- ISO 27001 — not certified
- PCI DSS — not in scope; do not put cardholder data in the workspace
- IRAP / Protected — not assessed
- “Sealed enclaves”, “100% immutable audit”, or “0 guardrail violations” — leftover prototype copy, not live product claims
2. Architecture, in one paragraph
A Member signs in through Base44. The browser calls one allowlisted proxy function. That function attaches server-side identity and talks to the Atlas bridge through a Cloudflare tunnel. The bridge is the only thing that talks to the database, private file storage, connectors, and Hermes. The browser never receives bridge tokens, database URLs, company IDs as an authorisation input, or model credentials.
3. Controls that are in place
- Server-derived identity and roles (owner, manager, staff). Browser-supplied role or company fields are not trusted.
- Invitation-only membership. The contact form cannot join a workspace.
- Company-scoped database access, with row-level security enabled on application tables and a non-bypass runtime database role.
- TLS required for the database connection.
- Private file storage on Atlas for uploads and task artifacts. Authenticated download. Size and type limits on library uploads (PDF, DOCX, XLSX, TXT, MD, CSV, 10 MiB).
- Mailbox and connector secrets encrypted at rest. Connector credentials stay on the bridge; agents see a loopback proxy, not the key.
- Company connectors are read-only and operator-installed. Personal Microsoft mail tokens cannot send or delete mail.
- Draft-first AI: structured drafts, human review, a separate approval gate, and a server executor. Unknown external actions fail closed.
- Ask AI history purged 30 days after a thread goes quiet.
- Library uploads stay untrusted until a manager or owner approves them for AI use.
- Public contact form: fixed recipient, honeypot, field limits, isolate-local throttle. Do not treat that throttle as edge-rate-limiting.
4. Remaining risks customers should know
- Model / inference credentials currently live in the agent environment that can run tools. A hostile prompt plus egress could theoretically reach those credentials. An inference broker outside the tool sandbox is planned, not done.
- Native learned skills can persist across tasks. Untrusted task input can influence later runs until learning is moved behind the same approval path as other writes.
- Agent network egress is not yet a strict destination allowlist. Review remains the control for what leaves in a draft.
- Markdown previews can request remote images unless a later change sandboxes that.
- There is no customer self-service export or deletion console yet. Those are operator actions.
- Contact-form throttling is local to the function isolate.
- Independent restoration tests from backups, and a full historical review of an earlier database-API exposure window, are operational follow-ups — not completed customer assurances.
A September 2026 internal security assessment closed a critical public-table exposure and several high-severity isolation issues. Passing regression tests is not a security certification. Customers with a low risk appetite should wait for the inference-broker and learned-skills work, or scope the first workflow so the AI never sees their most sensitive mail and files.
5. What we expect of the Customer
- Invite only people who should see the company’s workspace.
- Approve library files deliberately. Unapproved files are not AI context, but they still sit on disk.
- Do not connect a mailbox that contains material the company is not willing to have drafted from.
- Review drafts. The product is built around that step.
- Tell Lumevo promptly if a Member should lose access, or if you suspect a compromised account.
6. Incidents
Report suspected incidents to support@lumevo.digital with “security” in the subject. Notification to the Customer follows the Data Processing Addendum. If Lumevo is an APP entity and an eligible data breach occurs in Lumevo-controlled account data, Lumevo will assess and, where required, notify the OAIC and affected individuals under the Notifiable Data Breaches scheme.
7. Changes
Security changes that materially reduce protection will be treated as material contract changes under the Terms of Use. Improvements may ship without that notice.
