"AI Just Found 28 Robot Zero-Days in 48 Hours — Including One That Kills the Emergency Stop"

Here's a number worth sitting with: 28. That's how many zero-day vulnerabilities automotive cybersecurity firm VicOne says it surfaced in under 48 hours, using AI-assisted research to audit the software behind the "Physical AI" ecosystem — tools and projects developed or maintained by Google, NVIDIA, and Meta, among others, according to The Robot Report.

One of them is not theoretical. In a separate demonstration, VicOne's researchers exploited a flaw to disable a robot's remote-stop command — the last-resort safeguard that is supposed to halt a machine no matter what else goes wrong.

Now, the important grain of salt: this claim arrived as sponsored content, published by VicOne itself, ending with a pitch for its Radeis and Rthena security products and an upcoming white paper. The findings are explicitly "initial" — VicOne says the count may change after vendor and vulnerability-authority review, and some findings don't have CVE tracking numbers yet. As aiweekly noted, the post carries no named researcher quotes and is careful to call the wider consequences "potential rather than demonstrated outcomes."

But the shape of the findings is the story either way. Three quarters of the 28 sit in development, training, simulation, and validation tools — not in software running on deployed robots. A compromised model file or simulator could poison a robot's training data or falsify its test results, so it passes safety tests it should fail and ships with a badge it didn't earn. The attack doesn't have to reach the robot; it just has to reach the evidence that the robot is safe.

That inverts the usual security story. Robotics teams worry about the machine on the floor, but the fastest route to a bad robot increasingly runs through the laptop that trained it and the simulator that signed it off. When AI can audit that whole toolchain in two days — and attackers have the same AI tools — the patch window is measured in days, not release cycles.

What to watch: coordinated disclosure. When (and whether) Google, NVIDIA, and Meta acknowledge these findings with CVEs will tell us whether this was the future of security research, or a vendor using its own results as a brochure. Either way, the broader point stands: physical AI has a software supply chain, and it is being probed at machine speed now.

— via "The Robot Report"