One angle I haven’t seen much in red teaming writeups is “hardware-in-the-loop” failure evidence for embodied systems: not just whether the agent produces bad plans, but whether actuator-boundary constraints actually prevent out-of-bounds commands under malformed traffic / fuzzing. Curious if you’ve seen good frameworks for making those claims reproducible.
One angle I haven’t seen much in red teaming writeups is “hardware-in-the-loop” failure evidence for embodied systems: not just whether the agent produces bad plans, but whether actuator-boundary constraints actually prevent out-of-bounds commands under malformed traffic / fuzzing. Curious if you’ve seen good frameworks for making those claims reproducible.