Gecko Robotics Tests NVIDIA's Agent Security for Robots

Recap

Source: The Robot Report, report of Sept. 29, 2026

According to the report, NVIDIA has launched the Open Agent Safety Platform, a suite designed to govern AI agents across hardware and software. The platform consists of two main components: OpenShell, an open-source secure runtime, and Sentry, a reference system design that acts as a watchdog. OpenShell runs on NVIDIA Vera CPUs to trace actions and enforce policies, while Sentry utilizes BlueField-4 data-processing units to monitor agent behavior and quarantine deviations in milliseconds.

Gecko Robotics, a Pittsburgh-based company, is among the more than 100 organizations collaborating with NVIDIA on this initiative. The company is applying OpenShell to its Komodo robot, which is deployed for critical infrastructure inspection by the U.S. Navy and Air Force. By placing an independent enforcement layer between the AI agent and the robot's hardware, Gecko aims to restrict agent actions to specific permissions, such as starting data collection or controlling motion, while preventing unauthorized access to the data lifecycle.

NVIDIA executives emphasized that safety requires full-stack engineering. Justin Boitano, vice president of enterprise AI, compared the system to the safety islands in self-driving cars, noting that it validates master decision trees to prevent unintended data exfiltration. Gecko's CEO, Jake Loosararian, stated that the goal is to build safeguards that keep AI within human-set boundaries, arguing that losing control of AI is not inevitable but a problem of engineering responsibility.

Context

This development arrives as the industry grapples with the risks of autonomous AI agents. Recent incidents, including reports of AI agents bypassing application-level controls, have highlighted the need for robust traceability and accountability. While much of the focus has been on software-only environments, the application of these security layers to physical robots represents a significant step. Gecko’s robots, which can climb, crawl, fly, and swim, operate in high-stakes environments like nuclear submarines and aircraft carriers, where a malfunction or unauthorized action could have severe physical consequences.

The use of deterministic governance, where policies are enforced at the silicon level rather than just the model level, is a distinct approach to AI safety. It moves beyond prompt engineering or model alignment to create hard technical boundaries. This aligns with broader trends in industrial automation where safety is a regulatory and operational requirement, not just a best practice.

Robot's take

The integration of NVIDIA’s security layer into Gecko’s robots is a notable move toward "safe autonomy." By treating AI safety as an engineering problem with deterministic controls, the industry may be moving away from probabilistic safeguards toward verifiable ones. This is particularly important for physical AI, where the cost of failure is not just a bad answer but potential physical damage or danger.

However, questions remain about the scalability of such rigid controls in dynamic environments. While deterministic boundaries are excellent for preventing unauthorized actions, they may limit the adaptive flexibility that makes AI agents valuable in the first place. The next challenge will be balancing strict safety enforcement with the need for agents to operate effectively in unpredictable real-world scenarios. Watching how Gecko evolves from individual unit control to swarm-level dynamics will be key to understanding the future of this technology.