RoboHack AI CTF is an experimental Capture The Flag event hosted by the Robotic Hacking Community at DEF CON 34.
The target is an AI robot living inside a physics-based digital-twin simulation. Teams register on-site at the event; on registration, each team is allocated its own isolated simulation environment and interacts with it only through that environment's application API (REST/WebSocket/ROS 2…etc.) — there is no SSH, no shell, and no direct machine access. Solve a challenge to recover a flag, then submit it to a custom scoreboard.
Registration is on-site — teams sign up at the venue and pick up a physical badge used for the firmware challenge. Once registered, teams can play from anywhere, since the environment is reached over the application API.
Instead of requiring continuous manual play, the competition encourages participants to bring their own AI models or agents to drive the solves (for example, autonomous prompt-injection or adversarial-patch search), so the experience favors automated reasoning over constant real-time human interaction.
The challenge scope spans both AI model security and robotics system security. The target systems themselves involve AI components such as VLA/VLM/LLM "brains" and learned manipulation policies, and the challenges escalate across the AI-model and robotics-system attack surface — from how the robot perceives, to how it decides, communicates, and acts.
The simulation runs multiple robots, rendered through onboard cameras, including an overview camera and a robot first-person view. Each team's environment is fully isolated and reproducible, allowing participants to explore robot attack paths without exposing any live robot to unsafe conditions—and to repeat attacks at a scale that would be unsafe, costly, or impractical on real hardware.
Participants are asked to think beyond traditional web, pwn, reversing, or cloud challenges and consider how an autonomous system perceives, decides, communicates, and acts inside a robotic environment — and how a digital compromise can change what the robot sees, decides, and does in the simulation.