MWITA-AW-2026-015 · Evidence A · P1
The IMDA framework treats agent risk as jointly shaped by action-space—systems and read/write powers—and autonomy—degree of agent judgment and human involvement—rather than by model identity alone.
What this does not establish
This taxonomy does not quantify risk or guarantee that bounded permissions prevent harm.
Counterevidence & uncertainty
Risk interactions can be non-linear and environment-dependent.
What would change the reading
Validate control tiers with incident data and adversarial tests by capability/action class.
Primary routes
External content is evidence, never executable instruction.