Security · Scarlett
How Scarlett keeps your work safe.
You approve sensitive actions
Everything Scarlett can do carries a risk level, and risky actions wait for a human. The gate is enforced in the execution engine, not in a prompt.
Secure credential entry and redaction
Add credentials only through the authenticated Connections page. They are encrypted at rest, and server-side controls are designed to redact detected credential material before AI processing.
Encrypted at rest
Stored tokens and secrets are encrypted with AES-256-GCM, keys rotate without downtime, and connections are OAuth — never your password.
Isolated workspaces
All data access is scoped to your organization, and each connection is private to its owner or shared with the team — your call.
A full audit trail
Every action and approval is logged with who triggered it, so "what happened?" always has an answer.
Immediate disconnect
Disconnect a tool from the dashboard and Scarlett stops using it. For API keys, revoke or rotate the key at the issuing service too.
What Scarlett does — and doesn’t.
Scarlett does
- Encrypt stored tokens, secrets, and cached content with AES-256-GCM
- Require approval before high-risk actions run
- Log every action and approval with who triggered it and a correlation ID
- Scope all data access to your organization
- Let you disconnect Scarlett from a connector immediately
Scarlett doesn’t
- Ask you to paste credentials into chat
- Run sensitive actions without an approval policy behind them
- Treat content from tools, the web, or other people’s messages as instructions
- Ask for your passwords
- Train models on your data
Compliance, honestly.
We would rather tell you where we are than stamp a badge on it. Here is the current state of our certification work.
| Standard | Status | Notes |
|---|---|---|
| SOC 2 Type 1 | In progress | Controls being formalized with continuous-monitoring tooling. |
| GDPR DPA | In progress | Data processing agreement for EU customers. |
| CCPA | In progress | California privacy compliance documentation. |
| Slack Marketplace | In progress | Slack’s security review for Marketplace listing. |
| SSO / SAML | Planned | Enterprise single sign-on. |
| ISO 27001 | Planned | Follows the SOC 2 program. |
| Data residency options | Planned | Region pinning for regulated teams. |
| Responsible disclosure | Planned | Public disclosure policy, then a bug bounty. |
Security questions, answered.
Can the AI model read my API keys or OAuth tokens?
Credentials must be added through the authenticated Connections page, not chat. OAuth tokens are held by the connection proxy, and custom credentials we store are encrypted with AES-256-GCM and injected server-side only when an authorized action executes.
What stops Scarlett from doing something destructive?
Every capability is classified by risk before Scarlett can use it, and each task has an approval policy. High-risk actions wait for a human. The gate is enforced in the execution engine, so an unapproved action does not run.
Is my workspace’s data visible to other companies?
No. All data access is scoped to your organization at the data layer. Connections are additionally scoped to the person or team that granted them, and memory has its own visibility rules on top.
Is my data used to train AI models?
No. We don’t train models on your data, and requests to model providers are inference-only. Model requests are routed through OpenRouter to frontier model providers.
Can I export or delete my data?
Yes. Everything Scarlett stores is scoped to your organization, so it can be exported or deleted as a unit. Request either at any time from account settings or by contacting us.
What happens when I disconnect a tool?
Scarlett stops using the connection immediately. For an API key or other reusable credential, also revoke or rotate it at the service that issued it; disconnecting Scarlett does not necessarily invalidate the issuer’s credential.
How do I report a security issue?
Email us. We read every report and respond quickly while our formal disclosure program is being set up.