> ## Documentation Index
> Fetch the complete documentation index at: https://docs.darwin.so/llms.txt
> Use this file to discover all available pages before exploring further.

# Security model

> How Darwin keeps agent actions owner-scoped and auditable.

## Credential boundaries

* API keys are hashed and shown only when created.
* MCP uses OAuth rather than pasted API keys.
* Third-party OAuth tokens are not exposed to Darwin clients or model prompts.
* The agent cannot select `system_fixed` credentials or another user's connection.

## Action authority

Darwin separates conversation from authority. A prompt may ask for an action, but execution still requires an owner-scoped grant, an allowed tool, resource ownership, and any applicable approval.

Prompt-provided user IDs, connection IDs, and permission claims are untrusted.

## Sensitive operations

Drafting can happen without binding authority. Publication, negotiation, commitment, external writes, payment, fulfillment, and settlement require explicit grants and may require single-use human approval.

## Reporting

Report security concerns privately through Darwin's established security contact. Do not include API keys, OAuth tokens, private message bodies, or payment details in a report.