Security

Discovery without expanding your attack surface.

• Read-only connectors. GitHub App read permissions, AWS AssumeRole with list/describe only, Entra directory read scopes.
• No secret values. We list secret names and env var keys. GetSecretValue is explicitly denied in the CloudFormation template.
• No prompt content. Scoring uses structural fingerprints — never conversation payloads.
• Tenant isolation. Postgres RLS on org_id for every table. Connector tokens stay server-side.
• Explainable scores. Every likelihood comes with evidence chips. No black-box classifier.