Sanctuary AI — agentic threat model

Not certain from the listing — The proprietary Carbon AI control system likely utilizes vision-language-action (VLA) or reinforcement learning models. Primary threats include adversarial physical inputs (e.g., visual patches that trick robot perception) and model reprogramming.

Not certain from the listing — Data operations likely involve massive datasets of physical telemetry, spatial mapping, and imitation learning. Threats include training data poisoning (leading to systematic physical failures) and spatial data exfiltration.

Not certain from the listing — Carbon acts as the orchestration framework translating cognitive goals into physical actuator commands. Threats include command injection bypassing safety limits and manipulation of the robot's goal-planning logic.

Not certain from the listing — Deployment involves on-device edge compute on the physical humanoid robot, likely with cloud connectivity for updates. Threats include physical tampering, local privilege escalation, and insecure over-the-air (OTA) firmware updates.

Not certain from the listing — Real-time physical safety monitoring and sensor telemetry logging are critical but unspecified. Threats include sensor spoofing (e.g., LIDAR/camera blinding) and blind spots in safety-critical override guardrails.

Not certain from the listing — Compliance with industrial robotics safety standards (such as ISO 10218 or ISO/TS 15066) is expected but not detailed. Lack of transparent access controls and audit logging for physical actions poses a major compliance risk.

Not certain from the listing — In industrial settings, robots may interact with other automated systems or fleet management software. Threats include cascading failures across a robot fleet and unauthorized peer-to-peer command propagation.