AivexaNewsSearch
AI news for builders and product teamsChecked every hour

28 Zero-Day Findings in Under Two Days. Is Robotics Security Keeping Pace?

Collected Nov 16, 2026

VicOne, a Tokyo-headquartered cybersecurity company focused on vehicles and robots, reported that it identified an initial set of 28 zero-day findings in less than two days using AI-assisted research. The sweep covered software and tools supporting the Physical AI ecosystem, including projects developed or maintained by Google, NVIDIA and Meta, among others. All 28 were classified as zero-day vulnerabilities on the basis of initial validation, though VicOne notes the count may change after vendor and vulnerability-authority review, and that some findings do not yet carry CVE tracking numbers.

A zero-day, in the usual sense, is a weakness that was previously unknown and has no available fix at the time of discovery. That matters here because the affected software is not peripheral. Many of the tools are used to train, simulate, test and operate robots, so a flaw in them can shape the models a team builds, the test results it trusts, and the instructions ultimately sent to a machine.

VicOne's breakdown places most of the risk upstream of deployment. About 75% of the findings in this research set involved development, training, simulation or validation tools, rather than runtime systems on the robot itself. The company groups the findings across three stages of the robotics lifecycle: development and training; simulation and validation; and deployment and runtime.

At the development and training stage, the illustrative scenario is a team pulling a pretrained model to adapt for its robot. Loading a malicious file could give an attacker a foothold in the training system, from which training data or the model itself could be altered. A tampered model might pass routine checks and still misbehave when it later encounters a specific object, scene or command. In practice, this is the supply-chain pattern developers already know from package ecosystems, applied to model weights and training pipelines: the compromise arrives through a dependency the team trusts, not through a direct attack on the robot.

The simulation and validation stage carries a subtler danger. If an attacker gains access to a simulation environment by way of a compromised scene or robot model, they could alter simulated sensor readings, physical conditions or test results. The robot would then appear to pass tests that should have flagged unsafe behavior, and the team could approve it for real-world use on misleading evidence. This is worth pausing on for anyone who treats simulation as a safety gate: the gate is only as trustworthy as the integrity of the environment producing it.

At deployment and runtime, VicOne describes attackers changing files that influence motion planning, or sending malicious map data that crashes an operator's monitoring display. Compromising a remote-operation computer could also put the commands sent to the robot at risk. Depending on the system and its safeguards, the consequences could range from altered movement to operators losing visibility and being unable to respond quickly. VicOne frames these as potential consequences rather than demonstrated outcomes for every finding; actual impact depends on the attacker's access, the system's design, and whether other defenses stop a change from reaching physical behavior.

Separately, VicOne researchers demonstrated one concrete case involving a protective function: they exploited a vulnerability to stop a robot's remote-stop command from working. That example is the most pointed part of the report, because an emergency stop is exactly the control teams assume will work regardless of what else fails. It suggests that safety functions backed by software should be tested under the assumption that the supporting software has already been compromised.

VicOne also argues that patching is not the end of the process. Even after a maintainer ships a fix, a robotics team still has to find affected systems, check compatibility, and confirm the update preserves intended behavior. Earlier passing tests remain useful evidence, but a newly found weakness may probe conditions those tests never covered. The company offers three questions for response: where the affected software is used, including training machines, simulation environments and deployed robots; what the weakness could change, whether models, test results, operator visibility or robot behavior; and how protection is verified, confirming both that the fix closes the hole and that the updated system still behaves as intended. Where immediate updating is impractical, temporary measures such as restricting access to an affected service may be needed while a validated fix is prepared.

The broader point concerns tempo. AI can speed up parts of vulnerability discovery, since the work involves reading code, tracing how systems communicate and checking whether a suspected flaw is actually exploitable. The same capabilities are available to attackers, and when a flaw is a zero-day, faster discovery compresses the window teams have to react. Exposure does not by itself mean an attack happened, but it does force teams to determine whether they run the affected software and what an attacker could reach through it.

Why it matters: Robotics and Physical AI teams now face a class of vulnerability that lives in their training and simulation stack rather than on the robot, and roughly three-quarters of these findings landed there. Developers should expect faster, AI-driven disclosure cycles and treat model files, simulation assets and operator tooling as part of the attack surface, not just the robotics code itself. A likely practical consequence, in my reading, is growing pressure to apply software supply-chain practices such as provenance checks and integrity verification to model weights and simulation environments, and to re-test safety functions after any patch. The exact severity of these 28 findings is still unknown pending coordinated disclosure and vendor review, so teams should watch for identifiers and advisories rather than assume impact from the count alone.

Read at IEEE Spectrum · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

This article is brought to you by VicOne . VicOne identified an initial set of 28 zero-day findings in less than two days using AI-assisted research across software and tools supporting the Physical AI ecosystem. These include projects developed or maintained by Google, NVIDIA, and Meta, among others. Based on initial validation, all 28 findings are classified as zero-day vulnerabilities. The count may change following vendor and vulnerability-authority review, and some findings do not yet have CVE tracking numbers