Securing Physical AI: Lessons from Exein for GCC Infrastructure

In September 2026, the Italian cybersecurity firm Exein secured $270 million in a funding round that valued the company at $1.7 billion, making it Europe’s most valuable private cybersecurity startup. The investment, led by Headline with participation from major players like Goldman Sachs and the European Investment Bank Group, was not just for another endpoint security tool. It was a bet on securing “physical AI”: the growing class of intelligent systems that perceive their environment and act upon the physical world. This includes everything from industrial robots and autonomous logistics fleets to smart utility grids and drone inspection systems. For enterprise operators in the Gulf Cooperation Council (GCC), a region undergoing a massive buildout of AI-enhanced critical infrastructure, Exein’s strategic pivot from general Internet of Things (IoT) security to this new frontier is a critical market signal. It validates the immense operational risk inherent in connected physical assets and provides a model for how a sovereign, on-premise AI security layer can provide the necessary control and visibility for the region’s most important industries.

A New Attack Surface: From Connected Devices to Physical Action

Exein’s evolution mirrors a broader shift in the threat landscape. The initial wave of IoT security focused on protecting a flood of new, often insecure, connected devices from being compromised for botnets or data exfiltration. The challenge was scale and diversity. The emerging challenge of physical AI is different: it is about consequence. When software can directly control machinery, vehicles, or industrial processes, a compromise is no longer just a data breach. It can cause physical damage, disrupt production, or create significant safety incidents. This is the thesis that drove Exein’s significantly oversubscribed funding round and its subsequent partnership with Deutsche Telekom IoT, announced in October 2026, to secure robots, autonomous machines, and other physical AI applications.

Operationally, this new attack surface is complex and layered. It extends beyond typical application security to include:

• Device firmware and embedded operating systems
• Software libraries and AI model runtimes
• Sensors, controllers, and actuators
• Software update mechanisms and device management APIs
• Machine-to-machine communication protocols

Conventional security architectures, which often rely on perimeter firewalls or cloud-based agents, are poorly suited for this environment. They lack visibility into the low-level operations of specialized hardware and often cannot be deployed on legacy or resource-constrained industrial systems. A security failure in this context is not a database leak; it is a robot behaving unsafely, a chemical process receiving incorrect settings, a city’s power grid being manipulated, or an entire fleet of autonomous trucks being disabled. Exein’s approach, which aims to embed security closer to the device and its runtime environment, acknowledges that the point of failure has moved from the data center to the physical plant, port, or city street.

The GCC’s Compounding Operational Technology Challenge

The concentration of advanced industrial assets across the GCC makes the region uniquely exposed to these risks. Nations like Saudi Arabia, the United Arab Emirates, and Qatar operate a dense and interconnected fabric of critical infrastructure, including some of the world’s largest oil and gas facilities, petrochemical plants, desalination plants, ports, airports, and logistics hubs. Ambitious national visions are accelerating the adoption of AI and automation across these sectors, from autonomous mobility and smart city platforms in the UAE to advanced mining and utility management in Saudi Arabia and sophisticated LNG operations in Qatar.

This rapid modernization creates a uniquely challenging technology environment. State-of-the-art AI systems are being layered on top of, and integrated with, decades-old industrial control systems (ICS) and operational technology (OT). The result is a complex, mixed estate of legacy PLCs, proprietary SCADA systems, modern cloud platforms, third-party maintenance software, and a constant influx of new equipment from a global supply chain. A compromised AI-enabled predictive maintenance system could, in theory, trigger a shutdown or alter operating parameters on a legacy refining unit, creating a safety and continuity risk far greater than a conventional IT breach.

This heterogeneity means that security and asset information is dangerously fragmented. For an operator trying to secure a port, an oil field, or a smart city grid, there is no single source of truth. The data needed to assess risk is scattered across disparate systems, formats, and teams, creating critical blind spots that a physical AI attack could exploit.

Five Bottlenecks Crippling OT Security Operations

For security and operations teams on the ground, the challenge of securing physical AI manifests as a series of practical, daily bottlenecks. These are not theoretical problems; they are the workflow failures that prevent organizations from effectively managing risk in a converged IT and OT environment.

1. Fragmented Asset Truth

An organization cannot protect what it cannot see. In most industrial enterprises, the definitive inventory of devices, software, and configurations does not exist. Asset information is split across computerized maintenance management systems (CMMS), enterprise asset management (EAM) platforms, network monitoring tools, vendor portals, procurement records, and engineering spreadsheets. Operations teams cannot reliably determine which firmware versions are running on their PLCs, which libraries are included in a robot’s controller, or which AI models are deployed in their autonomous vehicle fleet. Without a unified, accurate asset inventory, vulnerability management becomes an exercise in guesswork.

2. Unstructured Maintenance Knowledge

Even when a vulnerability is identified, the knowledge required to fix it is often locked away in unstructured formats. Decades of operational wisdom—equipment manuals, standard operating procedures (SOPs), incident postmortems, engineering notes, and vendor security advisories—exist as PDFs, scanned documents, emails, and local file shares. Connecting a newly discovered CVE to the specific affected asset, its location, its operational criticality, and the approved remediation procedure is a manual, time-consuming process that often fails under pressure.

3. Legacy-to-Modern Integration Gaps

Many of the most critical systems in an industrial environment are the ones least equipped for modern security. Legacy PLCs, remote terminal units (RTUs), and proprietary SCADA systems were designed for reliability and isolation, not for a world of pervasive connectivity and dynamic threats. They often cannot run a conventional endpoint security agent, have limited logging capabilities, and communicate over specialized protocols that IT security tools do not understand. This leaves security teams reliant on passive network monitoring, which provides an incomplete picture of device state and behavior.

4. Configuration Drift and Third-Party Risk

Industrial environments are not static. Contractors, systems integrators, and remote maintenance teams regularly access critical systems. Emergency repairs and operational adjustments can lead to configuration changes that deviate from approved security baselines. These changes are often poorly documented and tracked, creating a persistent “configuration drift” that introduces new vulnerabilities. Temporary credentials, insecure remote access tools, and a lack of granular access control for third parties represent a significant and often unmonitored risk vector.

5. Evidence and Prioritization Overload

Security Operations Centers (SOCs) are frequently overwhelmed with a high volume of low-context alerts. Vulnerability scanners, log aggregators, and threat intelligence feeds generate a constant stream of data. Without deep operational context, analysts struggle to distinguish a truly critical threat—an exploitable vulnerability on a safety-critical system—from a low-risk finding on a non-essential device. This prioritization challenge wastes valuable analyst time and increases the risk that a significant event will be missed in the noise.

An Operator's Blueprint for a Sovereign AI Security Layer

Addressing these bottlenecks requires a new architecture, one that moves beyond siloed tools and embraces a unified, AI-driven approach to understanding and managing risk. Inspired by the market validation of Exein’s model, a regional blueprint for a sovereign AI security layer would not seek to replace human operators or grant AI autonomous control over industrial systems. Instead, it would empower operations and security teams by using AI to synthesize fragmented information and provide actionable, context-rich insights. The core of this architecture rests on combining device-level data with deep document intelligence inside a controlled, on-premise environment.

The key components would include:

• Comprehensive Data Ingestion: The system would ingest data from every available source: firmware images and software bills of materials (SBOMs) for analysis, real-time telemetry from OT and IoT devices, network traffic, and logs from existing security tools.
• Bilingual Document Intelligence: Crucially, the platform would ingest and understand the vast repository of unstructured operational knowledge. Using advanced document intelligence, it would process maintenance manuals, SOPs, incident reports, asset inventories, and compliance documents in both English and Arabic. This turns static, inaccessible documents into a queryable knowledge base.
• Private AI Agent Reasoning: On-premise AI agents would then reason over this unified dataset. These agents, running on secure, local hardware, would connect the dots. For example, an agent could correlate a new vulnerability notice with the asset inventory, identify the affected devices by cross-referencing firmware versions, consult the ingested maintenance manual to find the approved update procedure, and check incident reports to see if similar issues have occurred before.
• Human-in-the-Loop Control: The output is not an automated action, but a highly contextualized recommendation for a human operator. The AI would explain the vulnerability, assess its potential impact based on the asset’s role, propose a specific remediation plan grounded in approved SOPs, and generate the evidence needed for compliance and audit. All actions that modify configurations or apply patches would require explicit human approval, aligning with emerging best practices for AI agent safety, such as the principles of identity, least privilege, and tool verification reportedly found in China’s 2026 draft guidance on agent security.

This model is a direct answer to the operational bottlenecks. It creates a single source of asset truth, unlocks unstructured knowledge, and provides the context needed to prioritize alerts effectively. By deploying this capability on-premise, organizations like those building on the DHIA private AI platform can ensure that sensitive operational data never leaves their control.

What this means for operators in the Gulf

For CISOs, plant managers, and heads of operations in the Gulf, the rise of physical AI security is a call to action. It confirms that the risk associated with connected operational technology is now a board-level concern with significant financial and safety implications. Relying on a patchwork of legacy IT security tools and manual processes is no longer a viable strategy. The path forward involves building a unified, data-driven security capability tailored to the unique demands of critical infrastructure. This means prioritizing solutions that offer deep visibility into OT environments, can make sense of complex and unstructured operational documents, and, critically, operate within the region's stringent data sovereignty and residency requirements. The goal is not just to buy another security product, but to build an enduring capability for operational resilience.

Data Sovereignty: The Case for On-Premise AI

For GCC enterprises, the question of where AI processing occurs is not just a technical detail; it is a fundamental requirement of risk management and regulatory compliance. Regional data protection laws, including the UAE's Federal Personal Data Protection Law and Saudi Arabia’s Personal Data Protection Law (PDPL) as enforced by SDAIA, place strict controls on cross-border data transfers. While OT telemetry may not always contain “personal data” in the legal sense, it is invariably treated as commercially sensitive, confidential, or critical national infrastructure information under sectoral regulations and national security policies.

Sending detailed plant diagrams, network telemetry, asset vulnerabilities, and maintenance histories to a multi-tenant cloud AI service hosted in another jurisdiction is a non-starter for most critical infrastructure operators. It creates an unacceptable risk of exposing operational topologies, production patterns, and security weaknesses. A truly sovereign AI security architecture must therefore be built on a foundation of local control. This requires:

• On-premise or in-country deployment: All data processing and model inference must happen within the organization's own data centers or a designated in-country cloud.
• Customer-controlled data and models: Organizations must control their own vector databases, model weights, and encryption keys.
• Complete auditability: Every query, agent action, and recommendation must be logged for security review, incident response, and regulatory evidence.

This sovereign-by-design approach, central to platforms built for regional needs, ensures that the powerful capabilities of AI can be leveraged without compromising control over the nation’s most sensitive operational data.

Lessons from the Global Landscape

Developments in the open-source community and from industrial AI leaders in China provide further context. Recent benchmarks like ReproBench, which tested LLM agents on their ability to reproduce known IoT firmware vulnerabilities, highlight the current limitations of autonomous AI. The reported success rate of just 5.3% in reproducing real vulnerabilities underscores that AI agents are powerful analysis accelerators, not infallible autonomous hackers. This finding reinforces the need for human validation and an architecture built around operator empowerment, not replacement.

Meanwhile, China’s reported push to deploy thousands of industrial AI agents by 2027, combined with the availability of powerful open-weight models like Huawei’s openPangu-2.0, demonstrates the global momentum behind this technology. The open-weight model allows organizations to run and fine-tune models on their own hardware, providing a pathway for sovereign AI deployments. However, it also places the burden of governance—managing model provenance, dependencies, licensing, and security—squarely on the enterprise. It is a powerful enabler, but one that requires a mature approach to technology governance.

The takeaway

Exein’s landmark funding round is more than just a validation for one company; it is a clear signal that the market for securing physical, AI-driven systems is here. For GCC operators managing the complex and high-stakes world of critical infrastructure, this presents both a challenge and an opportunity. The old model of siloed IT security tools is inadequate for the new risk landscape. The winning strategy will be to build a sovereign AI security layer that fuses device telemetry with deep document intelligence on-premise, giving operators the unified context they need to defend the physical assets that power the regional economy.

Request a demo