Security
Your agent runs on its own machine.
An Idle Work agent is not a row in somebody else's shared database. Every client gets their own isolated machine, a container or a virtual machine, with its own filesystem, its own network namespace and its own credentials, on hardware we own and operate in the United States. Below is exactly how that is protected.
Where we stand
What is actually in place.
Not a row of logos. Each item is a control you can ask us to evidence.
- In place
Single-tenant isolation
One container or VM per client. Client machines cannot reach each other's data.
- In place
Encrypted in transit and at rest
TLS 1.2+ at the edge, WireGuard between machines, AES-256-GCM on disk.
- In place
Zero-retention AI provider
The provider we route to does not log, store or train on your prompts or responses.
- In place
PCI DSS Level 1 payments
Card details go straight to Stripe. We never see or store a card number.
- In progress
CASA AL1 assessment
The App Defense Alliance assessment behind Google API access. Underway, reviewed by an authorized lab, renewed every year.
Infrastructure
How it is actually built.
Isolation, not a shared tenant
Each client's agent runs on its own container or virtual machine with its own IP address, filesystem and network namespace. There is no shared agent runtime and no shared workspace. Portal records are separated by tenant id with row-level security on top. Nothing in the fleet is exposed directly to the public internet: every management path runs over a private Tailscale WireGuard mesh.
Encryption and credential handling
External traffic is TLS 1.2+ through Cloudflare. Machine-to-machine traffic runs inside a WireGuard mesh. The database sits on an encrypted ZFS pool (AES-256-GCM), passwords are bcrypt hashed at 12 rounds, and financial access tokens are encrypted with AES-256 before they are written. Service credentials live outside the application database, readable only by the service that needs them.
Access control and review
Least privilege by role, multi-factor authentication on every internal system, SSH by key only with password authentication disabled, and two-factor on the hypervisor console. Access rights, service accounts and third-party integrations are reviewed on a quarterly cycle, and access is revoked within four hours of someone leaving. Each agent box reports its own health, and a failure pages us on first detection rather than waiting for you to notice.
Your control
Nothing happens that you did not allow.
You approve what matters
A new agent starts narrow: it watches, drafts and reports. Actions that leave your business, such as sending mail, posting or writing to a connected system, are turned on one at a time and run behind an approval gate you configure. Raise or lower the bar per agent at any time.
Your data never trains a model
Not ours, and not our provider's. Prompts and responses are not logged, retained or used for training, and that is the written policy of the provider we route to, not an assumption. If you choose to connect your own AI account instead, that provider's policy applies and we tell you so before you connect it.
Export or delete, any time
Your business data is yours. Ask for an export and we produce one. Ask for deletion and we destroy the machine and purge its backups, and we confirm when it is done.
Scoped, revocable connections
Every integration is connected by you, through the provider's own OAuth screen, with the narrowest scope that does the job. Revoke it from the portal or from the provider and the agent loses that access immediately.
Subprocessors
Everyone else who touches anything.
The complete list. Each one does a single named job.
| Service | What it does | Their compliance |
|---|---|---|
| Ollama Cloud | AI model inference | Published policy: prompts and responses are never logged or used for training. |
| Cloudflare | DNS, TLS termination, edge protection | SOC 2 Type II, ISO 27001 |
| Stripe | Payments and subscriptions | PCI DSS Level 1, SOC 2 Type II |
| Plaid | Bank connections, where a client enables them | SOC 2 Type II, ISO 27001 |
| Google Workspace | Our own email and documents | SOC 2 Type II, ISO 27001 |
| Tailscale | Private network mesh | SOC 2 Type II |
Security questionnaire, architecture diagram or full policy pack: ask and we send it the same day.
FAQ
The questions procurement asks.
On hardware we own and operate in the United States, not on rented cloud capacity. The only exceptions are the named subprocessors above, each doing one specific job.