Miro AI Security

We operate under a conservative data-handling posture. Treat prompts and outputs as sensitive by default, and avoid unnecessary data retention.

Data Handling Rules

• PII redaction before logging. Opt-in only for debug sessions.
• Use allowlists for tools and APIs available to agents.
• Encrypt secrets end-to-end; never commit tokens.

Threat Model

• Prompt injection: constrain tool use and validate inputs strongly.
• Data exfiltration: outbound network calls are filtered and audited.
• Model supply chain: pinned versions and attestation for provider SDKs.

Expectations for Engineers

1. Default to least privilege for services and feature flags.
2. Route all model access through approved gateways.
3. Engage Security Engineering for novel data flows or third-party tools.