Securing AI Infrastructure Starts With Securing Trust

AI has moved rapidly from experimentation to operational reality. Boards, executive teams and technology leaders are under pressure to identify where AI can improve productivity, reduce costs and create competitive advantage. Yet as organisations integrate large language models (LLMs), AI agents and generative AI applications into business processes, many are discovering that security, governance and operational control have become just as important as innovation.

The pace of adoption highlights how quickly AI has become embedded in enterprise environments. According to the Fortinet 2026 Cybersecurity Skills Gap Global Research Report, 91 per cent of organisations are already using or experimenting with AI-powered cybersecurity solutions (1), underscoring the growing reliance on AI across business and security operations. At the same time, the report found that 50 per cent of organisations cite data privacy and information security as their biggest challenge when implementing AI (2), highlighting the growing need for stronger governance and security controls as adoption accelerates.

The conversation around AI is often dominated by its capabilities. Organisations want to know what AI can do, how quickly it can be deployed and what value it can deliver. However, the more important question is increasingly becoming: how do we secure it?

Cornelius Mare, Chief Information Security Officer, Australia, Fortinet, said, “AI adoption is accelerating because organisations recognise its potential to improve productivity, decision-making and operational efficiency. The challenge is that every new AI model, application and agent also introduces a new attack surface. Businesses need to think about governance, security and resilience from the outset rather than treating them as an afterthought.”

Many organisations focus on securing the applications that connect to AI models; however, the models themselves require protection. Unlike traditional applications, LLMs interact using natural language, creating risks that conventional security controls were not designed to address. This shift introduces a new set of security concerns that span governance, intellectual property protection, operational resilience and cost management.

The risks extend beyond external attackers. AI systems can be manipulated through malicious prompts, excessive consumption attacks, model poisoning or unauthorised access. Organisations must also consider how AI models interact with other systems, what data they access and what actions they are permitted to perform. Without appropriate controls, an AI application can become a pathway to sensitive information, critical infrastructure or business processes.

Cornelius Mare said, “Security leaders need visibility into who is accessing AI models, what those models are accessing and how information is moving through the environment. The goal is not to restrict innovation. It is to create the guardrails that allow organisations to innovate safely.”

The challenge is compounded by a growing shortage of specialist skills. According to Fortinet’s research, 60 per cent of organisations report difficulty finding cybersecurity professionals with AI-specific expertise, while 63 per cent expect to require additional AI oversight and governance roles within the next three years. (3) As AI becomes more deeply embedded in operations, organisations must have the people, processes and governance structures needed to manage these technologies effectively.

The emergence of model context protocol (MCP) servers, AI agents and interconnected AI ecosystems adds another layer of complexity. AI models are no longer operating in isolation. They increasingly interact with external tools, application programming interfaces (APIs), databases and internet-based services. Every connection creates potential exposure and expands the number of pathways available to threat actors.

This evolving landscape is reflected in guidance from the OWASP Top 10 for Large Language Model Applications, which highlights risks including prompt injection, model poisoning, excessive resource consumption and sensitive data disclosure. (4) These risks require security controls that can understand context and intent within natural language interactions rather than relying solely on traditional network or application-layer protections.

At the same time, organisations must maintain both security and performance. AI initiatives are often measured by their ability to deliver business outcomes at scale. Security controls cannot become a bottleneck that increases latency, disrupts user experience or drives up operational costs. Instead, security must be embedded into AI infrastructure in a way that supports performance, governance and compliance simultaneously.

For organisations operating AI workloads in cloud environments, visibility becomes even more important. Hybrid and multi-cloud architectures can create blind spots across data flows, permissions and workloads. Understanding how AI applications interact with cloud resources, where sensitive information resides and how access is managed remains fundamental to reducing risk.

Mare said, “The organisations that will gain the greatest value from AI are those that approach it as both a business opportunity and a security challenge. AI infrastructure needs the same level of governance, visibility and operational discipline that organisations apply to critical business systems.”

As AI becomes embedded across enterprise environments, securing the underlying infrastructure will become a core requirement for maintaining trust, resilience and business continuity. The organisations that succeed will be those that build security into every stage of the AI lifecycle, from model deployment and access management through to runtime protection, monitoring and compliance.

Mare said, “AI is changing how organisations operate. Securing AI infrastructure is about ensuring these systems remain trustworthy, resilient and aligned with business objectives so organisations can innovate with confidence as adoption continues to accelerate.”

Last Updated on July 14, 2026 by Nick Ross