Physical AI, the technology that lets robots perceive, decide, and act in the real world, is advancing rapidly, but a major blind spot threatens to undermine the entire field: security. Universities, defense agencies, and industry leaders are now convening to address what experts call the most pressing challenge in embodied AI development, as robots transition from controlled factory environments into homes, hospitals, and conflict zones (Source 1, 2).

What Exactly Are the Security Risks in Physical AI?

Physical AI systems differ fundamentally from traditional software because they operate in the physical world, making them vulnerable to attacks that go beyond typical cybersecurity threats. A robot's sensors can be spoofed, its actuators can be compromised, and its decision-making algorithms can be manipulated in ways that cause real-world harm. Unlike a hacked website, a compromised robot can cause physical injury or property damage.

On July 22, 2026, Carnegie Mellon University and the Carnegie Bosch Institute organized an industry panel titled "Security Challenges in Physical AI" that brought together 82 participants representing 78 countries, underscoring the global urgency of the issue. The panel featured leaders including Andrew Moore, CEO of Lovelace AI and former Dean of Carnegie Mellon's School of Computer Science, alongside executives from Bosch and local robotics startups.

"Security Challenges in Physical AI" was organized in partnership with the Carnegie Mellon Institute for Strategy and Technology as part of the U.S. Army War College International Fellows' visit to Carnegie Mellon University," noted Alessandro Oltramari, president of the Carnegie Bosch Institute.

Alessandro Oltramari, President of the Carnegie Bosch Institute

The timing is critical. Robots are already being deployed in sensitive real-world scenarios. The Chinese government has deployed Walker S2 humanoid robots from UBTech Robotics to work at border crossings with Vietnam, while both American and international military forces are increasingly relying on drones and autonomous systems in conflict zones. These deployments highlight why security cannot be an afterthought.

How Are Researchers Building Safer Embodied AI Systems?

• Neuro-Symbolic AI Integration: Carnegie Bosch Institute is backing a symposium on "Embodied Neuro-Symbolic AI for Reliable and Safe Robotics" scheduled for September 27, 2026, at IROS 2026, one of the largest robotics research conferences worldwide. This approach combines neural networks with symbolic reasoning to make robot decision-making more transparent and verifiable.
• Real-World Task Learning: The institute is supporting a National Science Foundation research proposal focused on enabling intelligent agents to learn complex manufacturing tasks based on real-world scenarios, led by Wright State University in partnership with Bosch Research and LAVORO, a startup founded by Jean Oh, associate research professor in Carnegie Mellon's Robotics Institute.
• Human-in-the-Loop Dialogue: If funded, the project will include adaptive learning through dialogue with humans, allowing robots to ask clarifying questions and verify instructions before executing critical tasks, reducing the risk of misinterpretation or malicious commands.

Beyond academic research, industry is also shifting its focus. The Carnegie Bosch Postdoctoral Fellowship program expanded this year from four to six fellows, with one cohort specifically focused on "contact-rich robotics," which involves robots that must safely interact with humans and objects in unpredictable environments. This expansion signals that institutions recognize the depth of work needed to make physical AI trustworthy at scale.

Why Is the Automotive Industry Suddenly Focused on Embodied AI?

The automotive sector is emerging as a critical testing ground for embodied AI security and reliability. As vehicles evolve from "smart vehicles" into "intelligent mobile terminals," they are becoming platforms for embodied AI applications ranging from autonomous driving to in-vehicle robotics. The convergence of large language models, embodied AI, smart hardware, and AI agents is pushing the industry's innovation boundaries further.

Automakers and supply chain companies are actively exploring embodied AI applications across multiple scenarios, including smart manufacturing, logistics, intelligent mobility, and human-machine collaboration. The global humanoid robot market is expected to reach shipments in the 100,000-unit range by 2026, marking rapid growth compared to 2024 and 2025. This explosion in deployment means that security and reliability standards must be established now, before billions of robots enter everyday environments.

For the automotive industry, the challenge is not just technological but organizational. Companies must master core technologies like multimodal perception, large model inference, real-time control, and task planning, while simultaneously overcoming industrialization challenges such as reliability, cost control, mass production, and scenario adaptation. The focus of corporate competition is gradually shifting from single-product capabilities to system-level intelligence, where security is a foundational requirement, not an add-on.

What Does the Broader Robotics Landscape Look Like Right Now?

The robotics ecosystem is expanding rapidly across multiple domains. Chinese company UBTech Robotics launched the U1, a series of humanoid robots designed for personal companionship, featuring lifelike silicone skin and emotional artificial intelligence, with prices ranging from approximately $17,650 to $145,000. Meanwhile, Google DeepMind announced an upgrade to its Gemini Robotics software system, designed to run "truly adaptable robots" through a vision-language-action model that converts vision and language input into motor control.

Foundation Robotics demonstrated a new robot hand based on mechanisms imitating biological tendons, while an international academic team created a different kind of robot hand that can function when detached. Duke University produced Argus, a nonhuman robot designed to move and exert force in any direction, equipped with multiple cameras pointing in different directions. These hardware innovations are advancing rapidly, but without corresponding security frameworks, they risk creating new vulnerabilities.

The regulatory environment is beginning to respond. The U.S. Federal Communications Commission added foreign-built robots and power inverters to its Covered List of products deemed to pose an unacceptable risk to national security, following earlier actions regarding unmanned aircraft systems. In contrast, the National Highway Traffic Safety Administration granted permission to Zoox, an Amazon company, to deploy self-driving taxis in one city, showing that regulatory approval is possible when safety standards are met.

The convergence of academic research, industry deployment, and regulatory scrutiny suggests that physical AI security is transitioning from a niche concern to a mainstream priority. As robots move from laboratories into homes, factories, hospitals, and conflict zones, the stakes for getting security right have never been higher.