PHYSICAL AI
When AI can act, evidence must follow the action.
Physical consequences extend the question beyond a model output. An examination may need to establish who authorized an action, what constrained execution, which system state was active and how override or recovery was evidenced.
ACTION BOUNDARY
Security and functional safety remain distinct.
ACTION BOUNDARY
Control the transition from decision to effect.
The examination follows the point where an AI decision can change an actuator, environment or other physical state. Authorization, policy gates, vetoes and override paths belong in the evidence record.
EXAMINATION QUESTIONS
Four records at the point of consequence.
The scope depends on the specific system and action pathway.
Authority
What authorized the action?
Boundary
What constrained, denied or vetoed execution?
State
Which version, configuration and operating state produced it?
Recovery
What records show containment, override or return to service?
FRAMEWORK CONTEXT
PAI-SF™ provides the security architecture.
The Physical AI Security Framework uses Kinetic Zero Trust principles across sensing, decision-making, authorization and actuation. Its published control text establishes the applicable version and scope.
DISTINCT INTERFACES
Security assurance does not replace functional safety.
Sector-specific approval, functional-safety obligations and operational authorization remain separate interfaces. A bounded examination cannot stand in for them without an explicitly established basis.
DISCUSS A SYSTEM
Bring one physical action path and its records.
A useful starting question identifies the consequential action, the authority boundary and the available telemetry or decision record.