When AI can find 28 zero-days in under 48 hours, the vulnerability disclosure calendar just became everyone's problem.
The Summary
- VicOne used AI-assisted research to uncover 28 zero-day vulnerabilities in under two days across robotics and Physical AI tools from Google, NVIDIA, Meta, and others
- 75% of findings hit development, training, and simulation tools — the systems that teach robots what to do before they move
- The window between discovery and patch just collapsed, and attackers have the same AI acceleration available to them
The Signal
The robotics security model assumes a certain pace. Vulnerabilities get discovered, disclosed responsibly, patched on a predictable timeline. VicOne just compressed discovery from months to hours using AI, and the implications ripple through every stage of the robot lifecycle.
The 28 zero-days span tools used to train models, run simulations, validate behavior, and control deployed systems. These aren't edge cases in obscure code. They're weaknesses in infrastructure from Google, NVIDIA, Meta — the platforms teams rely on to build autonomous systems at scale. A compromised training tool poisons the model. A simulation weakness validates bad behavior as safe. A runtime exploit turns deployed robots into remote-controlled liabilities.
"When the weakness is a zero-day, faster discovery can leave teams with little time to respond."
What makes this different from traditional software security:
- Training-stage compromise means the robot learns wrong, and you won't know until it acts wrong in production
- Simulation vulnerabilities let attackers manipulate test results, making dangerous behavior look safe
- Deployment exploits turn physical systems into attack vectors — not just data theft, but movement and force
The AI acceleration cuts both ways. VicOne used it for research and disclosure. Attackers can use the same techniques for exploitation before patches exist. The traditional disclosure timeline assumed humans needed weeks to find, verify, and weaponize vulnerabilities. AI collapses that assumption.
The Implication
Robotics teams need incident response plans that assume zero advance notice. If your stack includes tools from Google, NVIDIA, or Meta's robotics ecosystems, audit which components touch training data, simulation parameters, or runtime control. The patch window is shrinking to hours, not days.
This is what security looks like when agents build agents. The vulnerability surface expands with every new tool in the Physical AI stack, and discovery speed just outpaced most organizations' ability to respond. The question isn't whether your robotics pipeline has vulnerabilities. It's whether you can patch them faster than someone else can find them.