AI-assisted security research has found 28 previously unknown vulnerabilities in robotics software in less than two days, according to automotive cybersecurity firm VicOne — a result that shows how quickly AI is compressing the timeline for finding flaws in the systems behind physical machines.
The findings, published by The Robot Report on November 1, span software and tools supporting what the industry calls the "Physical AI" ecosystem, including projects developed or maintained by Google, NVIDIA and Meta. VicOne said all 28 findings were classified as zero-day vulnerabilities based on initial validation, while cautioning that the count may change following vendor and vulnerability-authority review, and that some findings do not yet have CVE tracking numbers.
The headline number is only part of the story. Where the flaws sit matters as much as how fast they were found — and most of them are not in robots rolling around factory floors today.
Most flaws hide before deployment
About 75% of the findings sit in development, training, simulation or validation tools rather than in software running on a deployed robot, VicOne said. That distinction cuts both ways. A flaw in a training or simulation tool does not directly let an attacker grab a robot's controls, but it can corrupt the very process used to prove a robot is safe — if the simulation platform is compromised, the safety testing built on it becomes untrustworthy.
VicOne's findings span three stages of the robotics lifecycle: development and training, simulation and validation, and deployment and runtime. Simulation tools such as NVIDIA Isaac Sim, widely used to test robot behaviour virtually before real-world deployment, illustrate the risk: a weakness in the test environment can silently undermine confidence in everything tested there.
A stop button that would not stop
In a separate demonstration, VicOne researchers exploited a vulnerability to prevent a robot's remote-stop command from working. The remote stop is a protective function — the kind of safeguard meant to halt a machine instantly when something goes wrong — so a flaw that can neutralise it is among the most consequential class of findings in the set.
VicOne was careful about the limits of its own claim. The company described the wider consequences as potential rather than demonstrated outcomes for every finding, and the original post carried no named researcher quotes — it was published as sponsored content, with a pitch for the company's upcoming white paper and its Radeis and Rthena products. Radeis is aimed at finding vulnerabilities and testing robot behaviour in simulation before deployment, while Rthena handles threat detection during operation.
Why speed changes the disclosure race
The deeper point is about pace. A zero-day — a flaw unknown to the vendor, with no fix available — traditionally took weeks of expert labour to uncover. If AI-assisted tooling can surface dozens in under 48 hours, defenders and vendors have far less time to respond between discovery and potential exploitation, and coordinated disclosure becomes a race against the same tooling in attackers' hands.
As details emerge through coordinated disclosure, robotics manufacturers will need to map each finding to its effect on development, deployment and robot behaviour. The stakes are rising with regulation too: VicOne points to the EU Cyber Resilience Act, the EU AI Act, the EU Machinery Regulation and the IEC 62443 standard as frameworks now demanding traceable evidence that robots are secure. For an industry racing to put AI-driven machines into factories, warehouses and homes, the VicOne result is a warning that the software around the robot needs as much scrutiny as the robot itself.
