Security and data handling
How Duun isolates runs, protects secrets, limits what agents can do, handles and deletes your data, and never trains on it.
On this page
Duun runs agents against your code, your tools and your secrets, so the defaults are conservative: every run is isolated, secrets never reach the model, agents ask before risky actions, and your data is never used for training. This page is for engineers and admins evaluating or configuring Duun. For our company-level practices, see Security at Duun.
Training on your data
“Duun never trains models on workspace data. Providers are contractually barred too.”
That covers code, prompts, transcripts, artifacts, memories and anything agents read through integrations. The Train on our data setting in Settings → Storage & privacy is off and stays off. Share anonymous usage analytics is separate: it sends screen views and timings, never code, prompts or transcripts, and you can turn it off.
With bring your own keys, model calls go to your own provider account under your agreement with that provider.
Isolation between runs
- Each run works in its own git worktree or sandbox, on its own branch. Parallel runs never touch each other’s files.
- Cloud sandboxes are disposable. They stop when idle and are deleted when you delete them; their disk, caches and recordings go with them.
- On every machine, the agent runs as an unprivileged user with a per-run config, never in your home folder.
- Port forwarding is off until someone turns it on for a port, is logged, and stops when the machine does.
Secrets
Secrets live in Settings → API keys & secrets.
- Encrypted at rest. “Agents get it as an environment variable; the value never appears in transcripts.” Values are masked in logs, tool output and the run terminal.
- Scoped. Workspace secrets are available to every project unless a project overrides them; project secrets are readable only by runs in that project. Who can use limits a secret to specific agents.
- Ask before each use. The run pauses with an ask in your inbox before the value is mounted.
- Rotation. Paste a new value and choose what happens to the old one: stop using it now, or keep it for running jobs for 1 hour or 24 hours. Running agents pick up the new value on their next command. Revoking the old value at the provider is up to you.
- In workflows,
{{secrets.NAME}}references are allowed only in a Send to URL step’s headers and body, and are shown masked.
When you publish to Explore, secret values are stripped and replaced with placeholders. See Publishing to Explore.
What agents may do
Agents act through layered permissions:
| Control | Where | Effect |
|---|---|---|
| Tools | Each agent | Read & search repo, Edit files, Run shell commands, Open pull requests, Network egress |
| Permission mode | Agent or factory | Ask for everything, Ask for writes, Autonomous in sandbox |
| Command rules | Settings → Action approvals | Allow, ask or deny by pattern; first match wins |
| Factory guardrails | Factory → Guardrails | Can only tighten workspace rules |
| Integration write access | Each connection | Off makes the integration read-only |
Merges follow the workflow’s Deliver step: Open PR only leaves merging to a person. Branch protection on your repo host still applies to Duun like any other app. See Permissions and approvals.
Access and roles
| Role | Can |
|---|---|
| Admin | Billing, members, secrets and every project |
| Billing | Plan, invoices, payment and credits; no code or secrets |
| Member | Start runs and approve in their projects |
| Viewer | Read runs and artifacts; can’t approve |
Members and Viewers see only the projects they have access to. Factories and workflows add their own roles (Owner, Admin or Editor, User). See Members and roles.
SSO, SCIM and audit log
These depend on your plan. Availability per plan is announced at launch.
| Feature | Plan |
|---|---|
| SSO with SAML or OIDC | Business and Enterprise (plan-dependent) |
| SCIM user provisioning | Enterprise (plan-dependent) |
| Audit log | Business and Enterprise (plan-dependent) |
| Data residency, dedicated region | Enterprise (plan-dependent) |
Until SSO is on, members sign in with their email or a supported identity provider, and admins can restrict joining to the company domain with Anyone at @domain can join.
Data retention, export and deletion
All in Settings → Storage & privacy:
- Transcript retention: keep transcripts forever, or delete transcripts and logs older than a number of days. Deletion runs nightly.
- Export workspace data: transcripts, artifacts, settings and audit log as a
.zip. Duun builds the archive in the background and emails you a download link that expires in 7 days. - Delete workspace: deletes every project, run, artifact, secret and connection after 7 days. Billing stops the same day. You can cancel within 7 days; after that it’s permanent.
Local caches on your Mac (artifacts, worktrees) can be cleared from the same screen and don’t affect cloud data.
Integrations and webhooks
- OAuth connections request read access by default; write access is a separate choice. Disconnecting removes access immediately.
- Outgoing webhooks are signed with HMAC-SHA256 in the
X-Duun-Signatureheader, with a timestamp to block replays. Verify every delivery; see Verifying signatures. - Incoming trigger URLs require a bearer token (
ftk_…) shown once at creation. Rotating it makes the old token stop working.
API keys
Duun API keys (duun_sk_…, Preview) are created by Admins, carry explicit scopes, are shown once, and can be rotated or revoked. Use them only from servers, never from a browser or mobile app. See API overview.
Report a vulnerability
Email security@duun.ai with steps to reproduce and the impact you observed. Please don’t test against other customers’ workspaces or run automated scans against production. We acknowledge reports and keep you updated while we fix them.
For questions about personal data, see the Privacy policy or email privacy@duun.ai.