Security at Tandem

• Encrypted

  TLS for every connection; OAuth tokens and secrets encrypted at rest with keys kept outside the database.

• You approve

  Anything that sends, books or changes something starts on “Ask first”. Autonomy is opt-in, per action.

• Audit log

  Every read, draft and action is logged with the agent, tool and time — visible to you, always.

• Undo

  Approved writes wait in an outbox for a short undo window, and many actions can be reversed afterwards.

• Kill switch

  Pause one agent, or every agent at once. Paused agents read and send nothing.

• Least privilege

  Each tool asks only for the scopes its features need, added one tool at a time. Disconnect deletes the tokens.

• In transit: HTTPS only, including calls to Google, Microsoft, Slack and our AI provider.
• At rest: OAuth access and refresh tokens are encrypted with application keys that never leave the server configuration and are never sent to the browser or written to logs.
• Isolation: every record belongs to one workspace, and every query is scoped to it automatically, with tests that fail on an unscoped query.
• Minimal storage: content is fetched when a task needs it; we keep only what the task needs to show its work. See retention.
• No training: your data never trains generalised AI models. Our AI provider processes it under terms that forbid training, with zero data retention where available.
• Deletion: disconnecting a tool revokes and deletes its tokens; deleting your workspace removes everything, including backups within 35 days.

• Four permission levels per agent and per action — read, draft, ask first, autonomous. New agents can never send without asking.
• Typed tools only: agents can call a fixed list of actions (for example “create draft”, “send approved email”) — never arbitrary web requests.
• Prompt-injection defence: text in emails, documents and messages is treated as data. It cannot raise an action’s permission level, add tools, or skip an approval.
• Outbox and undo: approved writes wait out an undo window before they run, with the compensating action recorded.
• Rate limits on outbound actions, and kill switches per agent, per workspace and globally.
• No remote content: agent output never loads external images or follows links on its own.

You can see every step an agent took for any task, and every action it took, in Activity.

• Staff do not read customer content. Access is allowed only with your permission, for security investigations or when the law requires — limited to named people and logged.
• Hosting and production access are protected by multi-factor authentication; secrets live in server configuration, not in code.
• Dependencies are kept up to date and checked for known vulnerabilities.

• Google: Tandem’s use of Google data follows the Google API Services User Data Policy, including the Limited Use requirements. Gmail access is subject to Google’s OAuth verification and an annual independent security assessment (CASA).
• Microsoft: publisher-verified app registration with delegated, least-privilege Graph permissions.
• Privacy law: GDPR and UK GDPR — data-processing agreements with every subprocessor, export and deletion built into the product. Details in the Privacy Policy.

Found a vulnerability? Email hello@softwebhub.com with “Security report” in the subject. We acknowledge reports within two working days, keep you updated, and will not take legal action against good-faith research that avoids privacy harm and service disruption.