Trust center
An AI employee touches your systems and your data. This page says exactly how Sam does that, and what we have committed to by when. Nothing here says "in progress" without a date.
Commitments
| Item | By | Status (updated 2026-10-13) |
|---|---|---|
| SOC 2 Type II report | 2027-09-30 | Observation window opens 2027-03-01 |
| ISO 27001 certification | 2027-11-30 | Scoping with auditor |
| EU region (data and model endpoints) | 2027-08-03 | Designed in from day one; launches with the public release |
| HIPAA controls and BAA | 2027-11-03 | Healthcare pack |
| Penetration test (external) | 2027-03-15 | Booked before the private alpha ends |
| Public bug bounty | 2027-11-03 | After SOC 2 Type II |
Identity and permissions
- Every action runs as a specific principal: a person, through the accounts they connected and the scopes they granted, or Sam itself, through a service identity an admin owns. There is no pooled workspace credential.
- Role-based access controls who may ask Sam to spend money, send external messages, change records, deploy code or read sensitive data classes.
- Data boundaries: what Sam learns in one team is not used to answer in another unless policy allows. Private conversations are read only when Sam is a participant and only for that conversation.
- Sensitive actions require approval; money movement above a threshold requires two people. Sam cannot approve its own work.
Credentials and secrets
- Credentials live in per-principal vaults and are injected only at the tool gateway. They never enter model context, logs, or the work log.
- Personal, health and payment data are detected and tokenized before model calls where a pack requires it; the model sees placeholders.
- One click revokes every credential a Sam holds and exports its record.
Execution
- Each task runs in its own isolated sandbox with a network egress allow-list for that task. Nothing persists between tasks except what the work log and the org memory keep on purpose.
- Browser and desktop automation run inside the same sandbox with per-site credentials from the vault.
The work log
- Append-only. Every model call (provider, model, tokens, cost), every tool call (target, operation, digest, reversibility), every approval, every data read and every artifact.
- Replay, diff and undo. Export to JSON, CSV, PDF and your SIEM. Retained for at least six months by default, longer if you configure it, with legal hold.
Data
- Your data is pinned to the region you choose (US or EU) and is never used to train models. Model calls are made under zero-retention terms with each provider.
- Deployment options: Sam's cloud, your own cloud account, or self-hosted with your choice of model.
- Subprocessor list, DPA, data-flow diagram and deletion procedures will be published here before the public launch.
Regulation
- Sam discloses that it is AI to everyone outside your organization, on every channel including the phone. This cannot be switched off (EU AI Act Article 50, US state chatbot-disclosure laws).
- EU pack: AI-literacy module with completion records (Article 4), DPIA template, record of processing export, works-council documentation for Germany, the Netherlands and France.
- HR pack: high-risk mode for recruiting with a human decision required and four-year record retention.
- Finance pack: dual control, separation of duties, archive-ready exports.
Report a vulnerability
Email security@hiresam.io. We acknowledge within two business days.