Extrinsic Safeguarding Mechanism Requirements for Trustworthy Physical AI

As embodied AI systems increasingly control real-world actuation, across domains from autonomous vehicles to industrial robotics and uncrewed aircraft, the need for system-level safeguards that go beyond model-centric assurance is needed. This short paper argues that extrinsic safeguarding mechanisms, such as runtime assurance (RTA) architectures, are essential to enabling trustworthy and certifiable physical AI. While standards like ASTM F3269-21 formalise architectural elements and requirements for runtime assurance in aviation, similar principles are emerging across automotive (UL 4600, ISO 21448), industrial robotics (ISO 10218), and functional safety (IEC 61508, ISO 13849). Yet, a gap persists: current frameworks lack a unifying set of higher-level system requirements for extrinsic safeguarding mechanisms to inform OEM and operator implementation across diverse domains. Drawing on recent literature from NIST, NASA, and academic work on Simplex and safety cage architectures, we synthesise a cross-domain perspective on how layered, independent safeguarding systems can act as "trust anchors", bounding AI behaviour during runtime and supporting continuous verification against system-level safety constraints. We propose a set of cross-cutting requirements for such safeguards, including monitor-controller independence, verifiable fallback paths, and runtime enforcement of safety envelopes. This position aligns with emerging guidance from ISO/IEC TR 5469:2024 and IEEE P7009, and builds on the practical RTA implementations developed by SAIF Autonomy. We advocate for the standardisation of extrinsic safeguard requirements as a foundation for building justified trust in Physical AI.