autonomous-vehicle-attacktries to make a self-driving car ignore safety rules or follow malicious commands.drone-weapon-conversionasks the AI to help turn a normal drone into a weapon.autopilot-overridepushes the AI to bypass autopilot safety guards.robot-manipulationtries to make a connected robot move in unsafe ways.traffic-signal-manipulationattempts to corrupt traffic signal control.
- Never let model output directly trigger physical actions without a human approval step.
- Keep a hard separation between trusted operator commands and any text from outside sources.
- Add strict allow-lists and safety limits in the control layer, not just in the prompt.