Security

As intelligence moves to the edge, securing decentralized devices against emerging threats is paramount. This group will focus on defining best practices and robust architectures for protecting edge AI models, data pipelines, and end-user privacy at the point of action.

Charter & Goals

1. Introduction

  • 1.1. Work Group Mission Statement: As physical AI moves from labs into robots, autonomous machines, and safety-critical systems operating in the real world, security can no longer be an afterthought bolted on at the network perimeter. The Physical AI Security Working Group exists to make security a foundational property of every AI system that senses, decides, and acts in the physical world. We bring together security researchers, hardware and silicon vendors, robotics builders, AI developers, and academia to define shared threat models, best practices, and reference architectures that keep physical AI systems trustworthy from the silicon up to the model — and resilient against attacks that arrive through sensors, firmware, runtime behavior, and the AI decision layer itself.
  • 1.2. Key Definitions/Terms
    • Physical AI: AI systems that perceive, reason about, and act upon the physical world through sensors and actuators — including robots, autonomous mobile robots (AMRs), drones, industrial control systems, medical devices, and agentic systems that trigger physical actions.
    • Attack Surface (Physical AI): The full set of entry points an adversary can exploit across a physical AI system — spanning the AI/model layer, the software and firmware runtime, the operating system and kernel, and the sensors and physical interfaces.
    • Runtime Enforcement: Security controls that observe and constrain system behavior while the device is operating, blocking anomalous or malicious actions in real time at the OS/kernel level rather than only scanning at build time.
    • Build-Time Firmware Analysis: Static inspection of firmware and software before deployment to identify vulnerabilities, generate a Software Bill of Materials (SBOM), and assess CVE/CVSS/EPSS exposure.
    • Visual / Sensor-Based Prompt Injection: An attack in which adversarial input delivered through a physical sensor channel (e.g., a camera capturing crafted visual content) manipulates the behavior of an AI system.
    • Threat Modeling: The structured process of identifying assets, entry points, adversaries, and mitigations for a given physical AI system or deployment.
    • SBOM (Software Bill of Materials): A formal inventory of software components and dependencies within a device, foundational to supply-chain security and regulatory compliance.
    • Regulatory Frameworks (in scope): EU Cyber Resilience Act (CRA), IEC 62443, UNECE R156, FDA premarket cybersecurity guidance, and the Radio Equipment Directive (RED).
  • 1.3. Purpose: Clearly define the work group’s objectives. Some ideas: Enhance collaboration, accelerate innovation, foster knowledge transfer, address real-world challenges (also on wiki).

This Working Group aims to establish a common foundation for securing physical AI systems across the edge ecosystem. Specifically, the group will:

  • Enhance collaboration between security vendors, silicon and hardware providers, robotics and device builders, AI developers, and academic researchers who each hold one piece of the physical AI security puzzle.
  • Accelerate innovation by developing shared threat models, reference architectures, and defensive patterns that members can adopt rather than reinventing independently.
  • Foster knowledge transfer through educational content that demystifies how attacks against physical AI work and how they can be mitigated at the model, runtime, firmware, and sensor layers.
  • Address real-world challenges including regulatory compliance, securing AI decision layers against sensor-borne manipulation, protecting resource-constrained devices, and managing security across large, heterogeneous device fleets over long lifecycles.
  • 1.4. Deliverables (choose and define which apply to the topic):
  • Educational Content:
    • Tutorials, webinars, and seminars on physical AI threat modeling, runtime enforcement, secure firmware practices, and defending AI decision layers against sensor-based attacks.
    • Whitepapers detailing the physical AI attack surface, defensive reference architectures, and mappings to regulatory frameworks.
  • Community Engagement Activities:
    • Roundtable discussions bringing together security researchers, robotics builders, silicon vendors, and academia to align on shared threat models.
    • Live demonstrations of representative attack scenarios against physical AI systems and the controls that defend against them.
  • Knowledge Dissemination:
    • Regular updates on newly disclosed vulnerabilities relevant to robotics and physical AI (e.g., responsible-disclosure case studies), tools, and defensive techniques.
    • A curated summary of key developments in physical AI security.
  • Recognition and Amplification:
    • Case studies and success stories showing how members have hardened real physical AI deployments.
    • A spotlight series on individuals and organizations advancing physical AI security.
  • Practical Resources:
    • Demonstrations and walkthroughs of defensive solutions across the stack — ie, build-time firmware analysis, runtime observability, and kernel-level runtime enforcement.
    • Open-access reference threat models and best-practice checklists for common physical AI system classes (robots/AMRs, agentic workflow control, protected models).
  • Future-Oriented Initiatives:
    • A dynamic, evolving definition of “physical AI security” and its threat landscape that keeps pace with advancing capabilities.
    • Strategic plans to attract new participants across the robotics, silicon, and AI communities.
  • Collaborative Publications:
    • Co-authored articles and research between academic and industry members on emerging physical AI attack classes and defenses.
    • Annual reviews summarizing the group’s impact and progress in the field.

2. Foundational Principles (to be followed by WG)

  • 2.1. Shared Vision & Goals:
    • The Working Group holds a shared vision that physical AI systems should be secure by design — trustworthy from the silicon through the operating system, firmware, and AI decision layers. We align research goals with the real needs of organizations deploying physical AI: managing security across large heterogeneous fleets, meeting emerging regulatory obligations, and defending against attacks that traditional IT security tools were never designed to catch. The group is committed to vendor-neutral, ecosystem-wide best practices rather than any single product or approach.
  • 2.2. Open Communication & Transparency:
    • The Working Group will hold regular open meetings with published agendas and minutes (with a designated AI note-taker), maintain an online forum for asynchronous discussion, and host periodic workshops. In keeping with the sensitive nature of security work, the group will adopt responsible-disclosure norms: vulnerability details shared within the group follow coordinated-disclosure timelines, and published materials focus on defensive guidance and remediation rather than weaponizable exploit detail.

3. Evaluation & Assessment

  • 3.1. Metrics for Success: 

The Working Group will track, at minimum:

  • Number of published deliverables per year (whitepapers, threat models, best-practice guides).
  • Active member organizations and breadth across the ecosystem (security, silicon, hardware, robotics/AI, academia).
  • Attendance and engagement across meetings, webinars, and workshops.
  • Adoption signals — downloads, citations, and references to the group’s reference threat models and best practices.
  • Number of collaborative/co-authored publications between industry and academic members.
  • Contributions mapped to regulatory readiness (e.g., resources supporting members’ compliance efforts).

Where possible, the group will assess broader impact by documenting cases where its outputs measurably improved the security posture of deployed physical AI systems.

Members

Channa Samynathan
Channa Samynathan Chair Vice President, Integrations and Partner Strategy Exein
Judith Furlong
Judith Furlong Co-Chair Senior Distinguished Engineer Dell Technologies

Meeting Minutes

No meeting minutes published yet.

Output

No output published yet.

Articles