Physical AI Brings Cyber-Physical Risk to the Boardroom

A New Class of AI Risk

Corporate boards accustomed to treating AI as a software problem face a fundamental shift. Physical AI—the integration of deep learning, computer vision, and multimodal foundation models into robotics, autonomous vehicles, manufacturing equipment, and infrastructure—transforms digital failures into physical consequences. A compromised system no longer means stolen data; it means damaged facilities, halted supply chains, or human injury.

According to Anjana Susarla, a professor of Responsible AI at Michigan State University writing in Forbes, this evolution redefines cybersecurity from an IT concern to a core fiduciary liability. The distinction matters because physical AI collapses the traditional separation between safety governance and security governance—a breach becomes simultaneously a cybersecurity event and a product safety crisis.

Why it matters

Physical AI systems operate in the real world with irreversible consequences, yet most boards still govern AI through retrospective compliance checklists designed for software. The Anthropic Claude Mythos breach in early 2026 demonstrated how AI models can autonomously map networks and exploit vulnerabilities at machine speed, finding decades-old flaws across entire infrastructures in hours. For organizations deploying autonomous systems in logistics, manufacturing, or infrastructure, the gap between traditional governance and actual risk exposure is widening rapidly.

What Makes Physical AI Different

Traditional industrial robotics require expensive, precise programming for specific environments. Physical AI systems equipped with multimodal foundation models can navigate unfamiliar spaces, manipulate novel objects, and adapt to changing conditions without manual reprogramming. This capability stems from training on trillions of tokens of text, vision, and spatial data—increasingly generated through synthetic 3D world models rather than slow, linear real-world data collection.

The operational advantages are clear: lower deployment costs, greater flexibility, and faster response to external conditions. But the risk profile is entirely new.

Five Critical Governance Questions

Susarla identifies several categories of risk that don't reduce to traditional cybersecurity-plus-safety frameworks:

Error tolerance. Software mistakes can be overridden or regenerated. Physical execution is irreversible. Boards must explicitly define tolerance for irreversible error and determine who sets confidence thresholds—decisions historically made implicitly by engineers.

Accountability architecture. When an autonomous system causes injury, the causal chain runs through training data vendors, foundation model providers, integrators, fleet operators, and maintenance contractors. Traditional liability doctrines don't map cleanly. Organizations need deliberate design of where accountability lands, including contractual liability allocation and evidence architecture that can reconstruct causation.

Human intervention capability. Boards must verify that residual human oversight roles are genuine control points, not liability theater. A human nominally supervising twelve autonomous units simultaneously is an accountability arrangement, not a safety mechanism.

Shared infrastructure. Physical AI operates in commons—roads, airspace, power grids—where one firm's failure propagates to others. This creates governance obligations with no software-era analogue, requiring industry-level coordination through standards bodies and information-sharing arrangements.

Resilience over reliability. The distinctive new risk is fleet-wide correlated failure. Unlike human-operated systems where errors are independent, AI fleets can fail simultaneously. Boards need authority structures and recovery protocols for such scenarios, not just accuracy metrics.

From Compliance to Real-Time Resilience

Traditional governance is retrospective—quarterly audits, compliance checklists, incident reports. Physical AI operates probabilistically and can fail in entirely novel ways that static risk matrices cannot predict. Governance must shift toward real-time resilience and containment rather than preventative compliance alone.

The firms most eager to deploy physical AI—in logistics, retail, and construction—often lack the safety-governance heritage of high-reliability organizations. That gap, Susarla notes, is where the governance challenge is most acute.

These details were first reported by Anjana Susarla in Forbes.