Security and compliance

• Inference can run entirely on customer-controlled hardware.
• Events, snapshots, and audit records can remain in a customer-controlled data directory.
• Cloud models and external integrations are optional; administrators choose which services to enable.

• Command roles are separated into viewer, operator, and admin.
• The MCP endpoint uses bearer-token authentication; command tools are disabled by default.
• High-impact actions should require human approval and run through existing site-control systems.

• Status transitions are written to an append-only event log.
• Critical events can retain timestamped image evidence.
• Agent tool calls, notifications, and operator actions can preserve source and time context.

• Sensitive configuration is supplied through environment variables and should not be committed to source control.
• External alert webhooks support HMAC verification.
• Inference, notification, and agent-tool failures are isolated from the primary monitoring loop.

We do not currently claim SOC 2, ISO 27001, HIPAA, or other third-party certification. If your procurement process requires specific controls, a DPA, data residency, or a security questionnaire, contact us for a deployment review.

Do not disclose vulnerabilities in a public issue. Send reproduction steps, impact, and contact details to security@tensordynamics.org.