VicOne found 28 zero-day vulnerabilities in less than two days. The company used AI-assisted research across software and tools that support the Physical AI ecosystem, and IEEE Spectrum reported the findings. Some of the affected projects were developed or maintained by Google, NVIDIA, and Meta.
VicOne classified all 28 findings as zero-day vulnerabilities, which means the flaws were previously unknown and had no available fix. The count may change after vendors and vulnerability authorities review the work, and some findings do not yet have CVE tracking numbers.
The findings spread across three stages of the robotics lifecycle. A team downloads a pretrained model to adapt for its robot, and loading a malicious file could give an attacker access to the training system, changing the training data or the model itself.
An attacker who imports a compromised scene or robot model could alter simulated sensor readings, physical conditions, or test results, so the team might approve a robot for real-world use on the basis of misleading evidence.
An attacker could also change files that influence motion planning, send malicious map data that crashes the operator’s monitoring display, or reach a remote-operation computer and put the instructions sent to the robot at risk. These are potential consequences, not demonstrated outcomes for every finding.
The actual effect depends on the attacker’s access, the system’s design, and the other safeguards in place.
The Cost of Every New Function
Every new function added to a program carries a cost1. Each new function is a new potential failure point.
The discipline moves behavior without changing it, so data lives with the functions that use it, and modules communicate minimally. Every name, boundary, and abstraction earns its place through clarity rather than cleverness, because the real bottleneck in software is comprehension, not computation. In the robotics pipeline, the new functions are the frameworks, simulators, and test harnesses that Google, NVIDIA, and Meta now build.
A test can pass while still missing the code it is meant to check. The remedy is to be “naturally skeptical” of a test that has never been broken and to inject a fault into the code deliberately before running it, so that the test fails1.
The lesson for robotics is that a test harness that has never been pushed is no guarantee that a model will behave correctly when it leaves the lab.
A manipulated model could behave normally during routine testing but respond incorrectly to a particular object, scene, or command after deployment.
The Hidden Failures of Complex Systems
Complex systems accumulate hidden failure modes that testing rarely catches. Their components “typically have high failure rates in all tasks” and still accomplish their objectives through statistical averaging across levels2. Most evolved complex systems operate in nonlinear, near-critical regimes, close to thresholds and tipping points. The reasoning applies directly to the robotics pipeline.
A test harness, a simulation, and a training dataset are all components with large and poorly understood repertoires of behavior, and each one carries failure that no checklist can name.
The article makes the same point from the other side. Weaknesses in the tools could influence the models the teams build, the test results they trust, and the instructions sent to the robots. VicOne’s separate demonstration exploited a vulnerability to prevent a robot’s remote-stop command from working.
That is why teams need to test whether safeguards remain effective when the software supporting them is compromised. The finding does not say an attack has occurred, only that the exposure exists and the response time is short.
The Shipping Pace Outruns the Fixing Pace
Deploying changes in complex systems should rely on proven best practices to minimize risk3. Feature flags let teams enable or disable functionality without deploying new code. A canary release limits the rollout to a small, controlled subset of production. Progressive rollouts gradually increase the scope.
Each practice exists to give a team time to observe before the change spreads. The robotics pipeline now moves faster than that. AI-assisted discovery shortens the search for weaknesses, and attackers could use similar capabilities. VicOne surfaced 28 of them in under two days, faster than any of those practices was built to absorb.
The article sets out two demands for robotics teams. The first is to respond as AI helps uncover weaknesses faster. The second is to assess how each finding could affect robot safety and operations.
The order matters. A team must first identify which systems use the affected software, then determine what the weakness could change, and then confirm that a fix does not alter intended behavior. Finding a flaw and fixing a flaw are different work, and the article says a patch still has to be checked for compatibility and for whether it preserves the behavior it was meant to change.
The Patch Rate as the New Security Metric
If controlled rollout becomes the norm, robotics tool vendors could automate feature flags and canary releases to push patches the way their own tools push new features. Google, NVIDIA, and Meta could find themselves judged less by how many vulnerabilities their tools expose and more by how quickly fixes reach the teams that depend on them. The likelier outcome is that patch velocity becomes the measure of security, so a vendor that ships slower than its own tools change will lose the trust of the people building robots.
The question is no longer whether AI can find flaws but whether the people who built the tools can fix them faster, and the burden will fall on the teams who did not write the code in the first place.
Sources
“28 Zero-Day Findings in Under Two Days. Is Robotics Security Keeping Pace?.” spectrum.ieee.org.
1 Fowler, Martin. Refactoring.
3 Harsanyi, Teiva. Working on Complex Systems.
2 Krakauer, David C., Murray Gell-Mann, Kenneth Arrow, W. Brian Arthur, John H. Holland, Richard Lewontin. Worlds Hidden in Plain Sight.

Leave a Reply