Securing Industrial and Physical AI in Modern Manufacturing
October 3, 2026
Giovanni Gaggero | October 2026
AI is becoming part of the production environment
Hannover Messe 2026 put industrial AI and physical AI firmly on the manufacturing agenda, with demonstrations and discussions focused on increasingly autonomous production. AI is becoming part of the machinery that makes, inspects and moves products. Robots and other connected equipment can use it to interpret their surroundings, adapt to changing conditions and carry out more complex tasks.[1]
Physical AI describes AI embedded in systems that sense and act in the real world. Sensors provide information about the environment, models interpret that information, and actuators enable the machine to respond. Industrial AI applies AI to industrial operations, using production data, predictive methods and machine learning to improve processes and support decisions. The two overlap when AI helps industrial machinery decide how to perform its work.[2]
The market outlook reflects the interest in this technology. One industry forecast values the global physical AI market at approximately $5.0 billion in 2025 and projects $82.8 billion by 2035, with a forecast annual growth rate of 32.8% for 2026–2035. Manufacturing and automotive applications accounted for 23.1% of the market in 2025.[3]
Deployment is already substantial. A major global warehouse operator has reported installing its millionth robot across a network of more than 300 facilities. Its fleet includes autonomous mobile robots designed to navigate in shared working areas alongside people.[4]
For manufacturers, this opens opportunities to improve productivity, introduce more flexible processes and develop new ways of operating, but it also brings cybersecurity into direct contact with the decisions that govern physical activity.
Greater autonomy creates a more complex attack surface
When robots depend on AI running on IT systems, edge platforms or cloud services, new communication paths cross the boundaries that traditionally separated enterprise networks from industrial control environments. Model updates, task instructions and sensor feedback must move between these domains. If those connections are too permissive or bypass existing controls, they can erode segmentation and allow a compromise in the IT environment to reach machinery on the factory floor.
This interconnection further exposes weaknesses in legacy industrial protocols. Many were designed for restricted environments and provide limited or no built-in authentication, message integrity or protection against replay. Once an attacker gains access to these communications, those weaknesses may enable device impersonation, altered commands or falsified sensor feedback.
Adversarial AI adds further opportunities to interfere with machine decisions. Poisoned training data, tampered models, backdoors and manipulated inputs can influence an AI system’s outputs. Where language models are involved, prompt injection can redirect the intended task. These attacks may result in inappropriate actions being requested through otherwise legitimate control interfaces.
A zero trust approach therefore becomes essential across the IT/OT boundary. A device, application or user should receive no automatic trust simply because it operates inside the plant network. Access should be explicitly authenticated, authorised and restricted to the resources and actions required, with ongoing assessment of risk.
For physical AI, that principle also needs to inform how operators assess operational data. Even physical measurements cannot be trusted by default. A reported position, speed, force or distance may have been altered, replayed or generated by a compromised sensor. Authenticating its source and protecting transmission do not establish that the reading accurately represents the physical process.
Establishing confidence requires countermeasures suited to the machine: monitored communications, checks on measurement freshness and plausibility, and corroboration through sufficiently independent observations. Comparing reported states with expected physical behaviour can help expose inconsistencies. Where legacy devices cannot support modern security mechanisms, compensating controls must restrict their exposure and monitor their activity. These measures should complement appropriately engineered safety functions.
Without this validation, operators and AI systems may act on a misleading view of the machine, with consequences ranging from defective production and downtime to equipment damage or hazardous movement. Visibility into actual robot behaviour becomes a necessary part of deciding which information can be relied upon.
Airfield and Watchfield bring visibility to actual robot behaviour
Securing these environments requires manufacturers to connect the protection of enterprise systems with the protection of machinery and AI components. Each layer contributes information that the others need.
At Airfield, our contribution centres on visibility at the industrial field layer. With Watchfield, we help organisations gain visibility into the real behaviour of the robot, making that operational perspective available as they assess the security of connected machinery.
For an AI-enabled production environment, knowing that a robot is present and connected is only the beginning. Manufacturers also need to understand what it is doing within the process.
That requires visibility into industrial communications and their operational meaning. Protocols such PROFINET or Modbus carry information between controllers and field devices. Monitoring needs to account for the relationship between these layers.[5]
Watchfield gives manufacturers visibility into the robot’s real behaviour. Airfield’s focus on fieldbus monitoring brings attention to the communications close to industrial devices, where digital activity can be examined in the context of machine operation.[8]
This perspective is particularly useful when an apparently legitimate exchange has an unexpected operational effect. Valid credentials and correctly formatted messages do not establish that a task is appropriate for the current production state. Behavioural evidence helps teams investigate that question.
It also supports collaboration between cybersecurity specialists and process engineers. Security teams can examine access and software events, while engineers interpret the corresponding machine activity. Together, they can assess deviations, investigate their causes and determine an appropriate response.
As robots acquire more autonomous capabilities, this operational understanding becomes increasingly valuable. The assessment needs to include what the machine actually does after receiving and processing information.
Connecting operational visibility with security response
Information from the production environment should inform the organisation’s security operations. Operational context helps analysts assess which incidents could affect a production line and what further investigation is needed. It also helps process engineers understand why activity elsewhere in the network may matter to their equipment.
Airfield’s work includes OT visibility and automation for security operations centre workflows.[8] Within the wider security process, the aim is to make information from industrial operations usable by the people responsible for investigating and responding to threats.
NIS2 and IEC 62443 reinforce the need for operational visibility
For European manufacturers within its scope, NIS2 establishes legal obligations through national implementing legislation. Article 21 addresses cybersecurity risk management, including incident handling, business continuity, supply-chain security, vulnerability handling and asset management. Article 23 sets reporting obligations for significant incidents, including an early warning within 24 hours and an incident notification within 72 hours of awareness. The applicable national legislation determines an organisation’s specific obligations.
Alongside these legal requirements, the IEC 62443 series provides standards for securing industrial automation and control systems. IEC 62443-3-3 specifies technical system requirements covering areas such as continuous monitoring.
Operational visibility helps manufacturers put these requirements into practice. Watchfield’s visibility into real robot behaviour can contribute operational evidence to this wider effort. Combined with organisational procedures, appropriate technical controls and documented assessments, it can support manufacturers as they address their NIS2 obligations and apply relevant IEC 62443 requirements.
Resilience must develop alongside autonomy
Industrial AI and physical AI offer manufacturers greater flexibility and the ability to automate increasingly demanding work. Realising that value requires cybersecurity to develop alongside the technology and its connections.
The challenge combines the growing interdependence of IT and OT with attacks directed at AI itself. Manufacturers need to understand how information reaches a machine, how decisions are constrained and how actual operation can be observed.
With Watchfield, Airfield helps bring the real behaviour of the robot into view. That visibility gives manufacturers a stronger basis for investigating events and managing the risks of more autonomous production.
Contact Airfield to learn how Watchfield can support visibility into your robotic operations.
Sources
[1] Hannover Messe 2026, Accelerating Physical AI and Robotics for Autonomous Manufacturing. https://www.hannovermesse.de/event/accelerating-physical-ai-and-robotics-for-autonomous-manufacturing/ses/99984
[2] Lee et al., 2026 Roadmap on Artificial Intelligence and Machine Learning for Smart Manufacturing. https://arxiv.org/abs/2605.00839
[3] Physical AI Market Size, Share and Trends, forecast 2026–2035, updated September 2026. https://www.acumenresearchandconsulting.com/physical-ai-market
[4] Operator announcement of the one-million-robot deployment milestone. https://www.aboutamazon.com/news/operations/amazon-million-robots-ai-foundation-model
[5] NIST, Guide to Operational Technology Security, SP 800-82 Rev. 3. https://doi.org/10.6028/NIST.SP.800-82r3
[6] NIST, Adversarial Machine Learning, AI 100-2e2025. https://doi.org/10.6028/NIST.AI.100-2e2025
[7] 2026 State of Industrial AI Report. https://www.cisco.com/site/us/en/solutions/networking/industrial-iot/industrial-networking-report/index.html
[8] Airfield Security, Watchfield and field-layer cybersecurity. https://www.airfieldsecurity.it/
[9] Directive (EU) 2022/2555, Articles 21 and 23. https://eur-lex.europa.eu/eli/dir/2022/2555/2022-12-27/eng/pdf