Cybersecurity firm VicOne says it identified 28 previously unknown, or zero-day, vulnerabilities in software underpinning the physical AI ecosystem in less than two days, using AI-assisted research to scan code, trace system communications and probe suspected weaknesses. The affected tools, used for robot training, simulation, testing and operation, include projects developed or maintained by Google, Nvidia and Meta, according to The Robot Report.
Roughly 75% of the findings involved development, training, simulation or validation tools rather than live production systems, VicOne said, though the company cautioned that final vulnerability counts may shift as vendors complete their own review. In a demonstration of what the flaws can mean in practice, VicOne researchers exploited one vulnerability to disable a robot's remote-stop command, the kind of safety feature an operator relies on to halt a machine in an emergency.
Why a software bug can become a safety problem
Robotics has long treated cybersecurity and functional safety as separate disciplines: one team worries about hackers, another worries about a robot arm swinging into a worker. VicOne's research argues those lines are blurring. The firm grouped its findings into three stages of the robot lifecycle: development and training, simulation and validation, and deployment and runtime. A compromised training tool could let an attacker quietly manipulate the data a robot learns from. A compromised simulation tool could produce false validation results, making an unsafe robot appear tested and ready. And a compromised runtime tool, as the remote-stop demonstration showed, can remove the safety net entirely once the robot is already working.
That framing matters because physical AI software increasingly sits on top of a stack of open-source and third-party components that most robotics companies did not build themselves and may not audit closely. A vulnerability discovered in a widely used simulation or training library does not stay contained to one company's product; it potentially touches every robot maker that built on the same foundation, which is why vulnerabilities tracing back to Google, Nvidia and Meta tooling carry weight well beyond those three companies.
A fast-moving threat matched by fast-moving discovery
VicOne, a Trend Micro-backed firm focused on physical AI and automotive cybersecurity, says the 28 findings add to a track record of more than 180 zero-days it has uncovered across automotive and robotics systems, with over 100 million new threat-intelligence signals processed monthly. The company's pitch is that AI-assisted vulnerability hunting can now move at roughly the same speed robotics companies are shipping new AI capabilities, rather than lagging behind by months the way traditional manual security audits often do.
That speed cuts both ways. The same AI-assisted techniques that let a security firm find two dozen flaws in 48 hours could, in less careful hands, let an attacker do the same. As humanoid robots and autonomous machines move from research labs onto factory floors and public sidewalks, the margin for a safety system to fail quietly, without anyone noticing until something goes wrong, keeps shrinking. VicOne said its full findings and remediation guidance will appear in an upcoming white paper, with vendors given time to patch before further technical details are published.