Hackers Already Cracked Robot AI at DEF CON. Now There's a Tool to Test Yours Before They Do.

Key points

• VicOne released a free cybersecurity testing extension for NVIDIA Isaac Sim, aimed at robotics developers.
• At DEF CON 34's Robotic Hacking Community, 16 teams recorded 25 successful solves across AI and robotics security challenges.
• The extension lets developers run attack scenarios against their own robot models inside a simulation before any physical deployment.
• VicOne, a Trend Micro subsidiary founded in 2022, says it has found more than 180 zero-day vulnerabilities, meaning previously unknown security flaws, across automotive and robotics systems.
• Nearly every core challenge at DEF CON 34 was solved at least once, showing that digital attacks can alter robot behaviour across multiple layers of a system.

Robots are moving out of research labs onto factory floors, hospital corridors, and public streets. Decades of software security lessons are now arriving in physical form, where a compromised instruction doesn't just crash an app; it can move a machine.

VicOne, a cybersecurity company that spun out of antivirus giant Trend Micro, has released a free tool designed to answer that question before a robot ships.

What does the tool actually do?

The extension plugs into NVIDIA Isaac Sim, a simulation platform that robotics engineers already use to design and test robots virtually before building physical prototypes. VicOne's add-on layers attack scenarios on top of that simulation, so developers can watch their robot's behaviour change under digital pressure. Think of it like a crash-test programme, but for cyberattacks instead of collisions.

Developers load their own robot model and intended task. The extension then runs curated attack scenarios drawn from real security research. Teams can test whether a VLM, a vision-language model (software that lets a robot interpret what its cameras see and decide what to do), holds up when it receives manipulated images or hijacked commands. If the robot drifts from its task, crosses a boundary it shouldn't, or makes an unsafe decision, the team sees it in simulation rather than in a warehouse. As we noted on 7 August in our look at AI-guided robot arms moving into live warehouse environments, the gap between controlled demos and real deployments is exactly where these risks bite hardest.

Where did the attack scenarios come from?

VicOne ran a competitive hacking event called a CTF, short for capture the flag, at DEF CON 34, the annual Las Vegas security conference where researchers probe systems for weaknesses. The company invited security professionals to attack two physical robots: the Reachy Mini and the LeKiwi platform.

Sixteen teams recorded 25 successful solves. Attackers found paths through AI models, the robot's visual perception system, cloud connections, and communications links between components. One advanced challenge involving firmware, the low-level software baked into hardware chips, went unsolved; everything else fell.

VicOne studied both the successful attacks and the failed attempts. Failed attempts still reveal what attackers assume, which control paths they tried, and which safeguards almost gave way. All of that fed into the scenarios now bundled into the free extension.

Should robotics teams be worried right now?

Yes, and the DEF CON results explain why. The attacks didn't all come from one direction. They succeeded across perception, communications, AI decision-making, and embedded hardware. A robot that passes every functional safety test can still behave dangerously if an attacker manipulates what it sees or what it's told to do.

VicOne CEO Max Cheng put it plainly: "Cyber-safety validation must become a standard part of robotics development."

For organisations that need more than the free extension offers, VicOne sells a fuller version of its Radeis platform covering deeper investigation, remediation planning, and compliance documentation.

What should readers watch for?

The risk here isn't abstract. An attacker who can feed false images to a robot's camera, or intercept the commands it receives, could cause it to act in ways its designers never intended. Ask the teams managing those robots whether cyber-safety testing is part of their pre-deployment checklist and covers the full system, not just individual components.

The broader pattern is worth tracking. Robots are scaling into warships, warehouses, and construction sites fast enough that security tooling is only now catching up. Tools like this one matter most when they arrive before the deployments do, not after.