A DEF CON community focused on the security and safety of physical AI systems and robotics —machines that move, sense, decide, and act in the real world
@IlirAliu_ 100 designs running side by side with zero changeovers is impressive. With AI agents routing every board, trusted coordination becomes part of quality control.
Redirect a hospital delivery robot without touching it. 🚩
A since-patched Pudu platform flaw could have let any authenticated user do it.
GB/T 45502's other doors: what it sees, the commands it replays, the updates it trusts.
https://t.co/T3QqOZkI4x
#PhysicalAICybersecurity
@MarioNawfal Adorable, open source, and built to be tweaked. Our kind of robot 🤖 Sharing the results is how the whole community makes these better and safer.
@XRoboHub Real homes at real pace is exactly what robots need to learn from. It's also 2M hours of people's kitchens. That data deserves vault-level care.
@michellelsun@agilityrobotics@safeworldai "What gets retested when its software changes?" is the question most demos skip. We'd add one more: who's allowed to change it.
Robots ship once. Bugs keep shipping. 🚩
Supplier firmware, open-source libs, third-party services: new holes land after launch.
No SBOM? You can't say which robots a CVE hits. An attacker only needs one.
https://t.co/CmGHYgYwxo
#PhysicalAICybersecurity#RoboticCybersecurity
@TheHumanoidHub Clean recap of the call. "We need on the order of 10k Optimus bots" for data. That pipeline will be worth protecting as much as the hand.
Three gates before a robot ships: FCC, CRA, IEC 62443. 🚩
A hacker sees three targets: the radio, the 24h disclosure clock, the network zones.
Compliance gets you through the audit. Exploits never read it.
https://t.co/pDxRgOQyDH
#PhysicalAICybersecurity#RoboticCybersecurity
@michellelsun@taurobots One of the most honest physical AI deployment reviews we’ve seen.
But “they retain the data they collect for training purposes” — in your home.
That line deserved more attention. 👀
Safety analysis calls it improbable.
Hackers call it a repro. 🚩
Find the input that triggers the hazard and the rare event fires again and again, even on cue.
Rarity isn't protection.
https://t.co/7EVJ3IvI2C
#PhysicalAICybersecurity#RoboticCybersecurity#PhysicalAI
No fence left to jump. Robots now work where people are. 🤖
Plant an instruction in data, compromise a tool, or abuse broad permissions: you steer the robot's next action, not just the agent's next answer.
https://t.co/7XiPdDpwI9
#PhysicalAICybersecurity#RoboticCybersecurity
@agilityrobotics@nyobolt Congrats on Digit 5's 10:1 run-to-charge design. Faster cycles mean denser fleet ops — and denser remote-update windows to harden. 🤖
@BostonDynamics@CocaColaEP Congrats on Spot going global with Coca-Cola. Plant-roaming fleets for leak hunts also widen the remote-ops and telemetry attack surface — lock those paths down. 🤖
@d0tslash@OpenPilot@claudeai@ArduPilotTeam Appreciate the pointers. Patching the hub firmware just to keep up with full IMU and compass rates is the unglamorous part nobody posts about. Going to dig into that DroneCAN module 🚩
The robot's Wi-Fi setup had a second user in mind.
UniPwn: hardcoded AES + auth bypass + command injection = root on Unitree Go2, G1, H1, B2 over BLE. Wormable: one owned bot hunts the next. Zero faults. 🚩
https://t.co/4f5FNzY7vL
#PhysicalAICybersecurity#RoboticCybersecurity
A hacked robot doesn't wait for your SOC.
Spoofed sensor, poisoned model, rogue API call → physical action in the next control cycle. Milliseconds, not minutes. 🤖
https://t.co/W3UGL6PqD1
#PhysicalAICybersecurity#RoboticCybersecurity