VicOne announces it has identified 28 zero-day vulnerabilities in less than two days, thanks to AI-assisted research, in software and tools supporting the physical AI ecosystem. Among the affected projects are those developed or maintained by Google, NVIDIA, and Meta. This article is sponsored content produced with VicOne.
A zero-day vulnerability is a previously unknown flaw for which no patch is available. According to an initial validation, all 28 discoveries fall into this category. This count may change after review by the publishers and the vulnerability authorities, and some discoveries do not yet have a CVE tracking number, the standard identifier assigned to a vulnerability.
Many of these tools help teams train, simulate, test, and operate robots. A weakness in one of them could therefore influence the models teams design, the test results they rely on, and the instructions sent to robots.
Vulnerabilities discovered faster
Discovering vulnerabilities involves examining code, tracking how systems communicate, and verifying if a suspected weakness can actually be exploited. According to VicOne, AI can speed up part of this work and allow researchers to analyze more software in less time. Attackers, however, could have similar capabilities.
When a vulnerability is a zero-day, a faster discovery can leave little time to react. This exposure does not necessarily mean an attack has occurred, but teams must determine if they are using the affected software and what an attacker could view or modify.
Risks at every stage of the lifecycle
VicOne’s discoveries concern three stages of a robot’s lifecycle: development and training, simulation and validation, then deployment and operation. According to the company, 75% of the discoveries involve development, training, simulation, or validation tools, meaning risks can appear even before the robot is deployed.
During training, a pre-trained model, meaning one already learned on large amounts of data, can be modified without the team’s knowledge. It could then behave normally during usual tests but react incorrectly to a specific object, scene, or command after deployment. During simulation, a compromised scene or robot model could alter simulated sensor data or test results, leading to approving a robot based on misleading evidence.
During operation, an attacker could modify files influencing motion planning or send malicious mapping data capable of crashing the operator’s supervision screen. According to VicOne, compromising a teleoperation computer could also endanger the instructions sent to the robot. VicOne specifies these are potential consequences, not demonstrated results for each discovery.
In a separate demonstration, VicOne researchers exploited a vulnerability to prevent the remote stop command of a robot from functioning. According to the company, this case shows that teams must verify that their safety devices remain effective when the software that operates them is compromised.
Verify a patch before deploying it
Even when a publisher releases a patch, a team must identify the affected systems, verify the update’s compatibility, and ensure it preserves the expected behavior. A previously successful test remains useful, but a recently discovered vulnerability can reveal conditions that test did not examine.
VicOne proposes three questions to guide the response: where is the affected software used, including on training computers, in simulation environments, and on deployed robots? What could the vulnerability modify, whether it’s models, test results, operator visibility, or robot behavior? How to verify that protection is effective? When an immediate update is impossible, interim measures, such as restricting access to an affected service, may be necessary while awaiting a validated patch.
VicOne also presents two products: Radeis, which helps identify potential zero-day vulnerabilities, assess the security of AI models and systems, and simulate the effect of these risks on a robot’s behavior before its deployment; and Rthena, which provides threat detection and response within approved policies during operation.
