Security testing
We test Hivelock ourselves, using the open-source penetration testing tool Strix. Our latest test found 1 high-severity issue, which we fixed and retested.
Read the full security test report
These tests were run by our own team on a self-hosted test copy. Hosted billing was not part of the live test. An independent third-party penetration test is planned, and Hivelock does not yet hold a certification such as SOC 2 or ISO 27001.
Keeping customers apart
- Enforced in the database. PostgreSQL row-level security limits every query for workspace data to one workspace.
- Limited access. The application connects with a restricted database account.
- Checked on every request. Each workspace route checks the person's role. Someone outside a workspace gets "Not found".
- Separate keys. Each workspace has its own encryption key, so its encrypted data cannot be opened with another's key.
On the hosted service, customers share the same application and database, separated by these controls. If you need full separation, you can self-host.
Signing in
- Two-step sign-in. Every password account needs an authenticator app. It cannot be turned off, and devices are not remembered.
- Single sign-on. Connect any OpenID Connect provider, after proving your domain through DNS. Your provider then handles the second step.
- Roles. Owner, admin, builder, operator and viewer are checked on the server for every request.
- Sessions. They expire after about 8 hours without use, and a password reset ends other sessions.
- Attack protection. Sign-in attempts are rate limited, and the app uses a strict content security policy, CSRF checks and HTTPS.
Encryption
Traffic is protected with TLS. Hivelock also encrypts sensitive data itself with AES-256-GCM, using a separate key for each workspace.
| Encrypted by Hivelock | Use disk encryption for these |
|---|---|
| API keys, connection credentials and other secrets | Uploaded files |
| Workflow run data, chat messages and approval details | The document search index |
| Agent memory and document passages | Workflow definitions and agent instructions |
| WhatsApp contact numbers | Names of workspaces, people, files and workflows |
Controlling agents
- Only the tools you choose. An agent gets only the actions selected for it, and the model never sees your credentials.
- People approve changes. In chat, an agent asks before it changes data. No answer within 15 minutes counts as declined.
- Approval is by role, not by a second person. Anyone whose role can approve may approve their own run. Hivelock does not enforce separation of duties.
- Chats are shared within a workspace. Every member who can read workspace content, viewers included, can read all of its chats, including chats other people started.
- Unattended limits. An agent step in a workflow cannot use tools that need approval. A connector step you add runs as built, so put an approval step before it where needed.
- Reviewed tools. New or changed MCP tools stay off until someone reviews them.
- Sensitive-data removal. Common credential formats are removed from agent output and activity records. Personal-data removal is a workspace setting.
Prompt injection is an industry-wide risk. Hivelock marks outside text as data, limits tools and asks people before changes. These reduce the risk, but a model can still be misled, so give agents the least access they need.
Network protection
- Connectors cannot reach private or internal addresses, including through redirects or changed DNS answers. A self-hosted operator can allow named internal hosts.
- On a self-hosted server, owners and admins can point AI models at any address, including internal ones. Cloud metadata addresses stay refused. The hosted service cannot reach private addresses.
- Each credential is limited to its own service, so it cannot be sent elsewhere.
- Incoming webhooks must carry a valid signature, and replays are refused.
Billing
- No card details. Payment happens on Stripe's own pages, and Hivelock never sees card details.
- Signed events only. A plan changes only on a message signed by Stripe, checked over the exact bytes received and refused if more than five minutes old.
- Once and in order. Each event applies once, and a late event cannot undo a newer one.
- Checked nightly. The hosted service compares each subscription with Stripe, corrects any difference and records it in the audit log.
Audit trail
Hivelock records sign-ins, permission changes, approvals, runs, exports and deletions. The log is append-only and hash-chained, so a changed record can be detected. Each workspace can view its recent entries on the Audit page.
Your data
| Data | How long we keep it |
|---|---|
| Finished runs | 90 days by default. You can set 1 to 3,650 days. |
| Chats, files and agent memory | Until you delete them or the workspace is erased |
| Deleted hosted workspace | Access stops at once, and data is erased after 30 days |
| Audit log | Kept after erasure, because it cannot be changed. It holds ids, actions and client addresses. |
Owners of hosted workspaces can export their data. The export lists file names, but does not include file contents. Your content is not used to train AI models by Hivelock. Your chosen AI provider sees what agents send it, under that provider's terms. For the most sensitive work, use confidential mode or a local model.
What you control
- Choose AI providers whose data terms suit your work.
- Give connections and agents the least access they need.
- Keep approval steps in front of actions that change outside systems.
- When you self-host, protect the server, encryption keys and backups, and use disk encryption.
Report a vulnerability
Email sales@shichifukutekx.ae with what you found and how to reproduce it. Please do not include real customer data, and give us reasonable time to fix it before making it public. Read how we handle reports.