AI security: navigating threats, governance, and strategies

As artificial intelligence accelerates, businesses face a growing challenge: securing systems that may be exploited faster than vulnerabilities can be identified and fixed.

Jonathan Herman, CEO of SOVAIR, says organisations must rethink their approach to AI security as businesses increasingly deploy AI agents across sensitive operations, from handling confidential information to making decisions on behalf of employees.

Speaking with Brad Gastwirth on The AI Edge, Herman examined the changing cyber threat landscape, the risks of giving AI tools access to corporate systems and why businesses should consider stronger infrastructure protections as they prepare for 2027.

While concerns about rogue AI systems acting independently continue to attract attention, the more immediate danger lies in human actors using AI to launch attacks at greater speed and scale.

As AI tools become more capable, cybercriminals could use them to identify weaknesses, automate parts of an attack and exploit vulnerabilities before organisations have time to respond.

For businesses, that creates a growing gap between the speed of technological development and the ability of security teams to keep systems protected.

Why traditional security measures may not be enough

The growing use of AI agents is introducing new risks for businesses. Employees are increasingly using AI tools to process information, automate workflows and interact with systems containing sensitive corporate data.

Without appropriate safeguards, confidential information could be exposed through everyday AI use, while agents with excessive permissions could take actions beyond their intended responsibilities.

The challenge is particularly significant when AI systems have access to emails, financial platforms, customer records and internal databases.

Businesses must establish clear boundaries around what these tools can access, what actions they can perform and how their activities are monitored.

Herman’s advice highlights the importance of proactive security planning, rather than assuming existing controls will remain sufficient as AI capabilities advance.

However, restricting access is only part of the solution.

Organisations also need clear governance policies and effective communication with employees so staff understand the risks associated with sharing information through AI tools.

Rather than imposing blanket bans on advanced models, businesses should focus on educating employees about responsible use and establishing clear rules for handling confidential data.

Air-gapped AI emerges as a security consideration

One approach gaining attention in sensitive environments is air-gapped infrastructure, which separates computer systems from external networks, including the public internet.

Under this model, organisations can deploy AI models on-premises within a controlled environment, reducing exposure to remote attacks and limiting opportunities for unauthorised external access.

For businesses operating in critical infrastructure and other highly sensitive sectors, this approach offers an additional layer of protection by reducing the number of pathways attackers can use to reach AI systems.

Air-gapped systems are particularly relevant where AI interacts with sensitive operational data or supports functions that cannot tolerate significant security failures.

The approach is not without limitations. Isolating a system from the internet does not eliminate every security risk. Insider threats, compromised software, malicious updates and vulnerabilities within the isolated environment can still pose problems.

Nevertheless, Herman’s position points to a broader question for business leaders: whether keeping AI systems connected to external networks is always necessary, particularly when they handle sensitive information.

Simple sandboxing, which restricts an application’s access to other systems while potentially retaining internet connectivity, may not provide the same level of network isolation.

For organisations preparing their AI strategies for 2027, the decision should depend on the sensitivity of their data, operational requirements and the consequences of a potential breach.

Critical infrastructure raises the stakes

The implications extend beyond corporate data and commercial operations.

As AI becomes more deeply integrated into power grids, telecommunications, transportation and other essential services, securing these systems becomes increasingly important to operational resilience and national security.

In these environments, a security failure could have consequences beyond financial losses or the exposure of confidential information.

Organisations must therefore consider the risks associated with connecting AI systems to operational technology, while carefully controlling access and limiting the actions automated systems can perform.

For highly sensitive applications, physical network separation may form part of a broader security strategy alongside access controls, monitoring, testing and incident response planning.

Balancing AI productivity with business security

The challenge for executives is to capture the productivity benefits of AI without creating new vulnerabilities or undermining employee confidence.

As adoption expands, security cannot remain a final checkpoint added after a technology has been deployed.

It must be considered during the design, implementation and ongoing management of AI systems.

That means establishing clear governance frameworks, defining acceptable use, limiting data access and ensuring employees understand how their use of AI affects the organisation.

Businesses must also consider how AI-driven automation affects jobs and responsibilities, communicating changes clearly rather than allowing uncertainty to undermine trust.

The central issue is not whether organisations should adopt AI, but how they can do so while maintaining control over their systems, information and operations.

As AI capabilities continue to evolve, the organisations best positioned to benefit will be those that treat security as a core part of their AI strategy rather than an afterthought.

Herman’s insights underline the need for businesses to move beyond reactive defences and make deliberate decisions about where AI operates, what it can access and how its use is governed.

With 2027 approaching, those decisions could become increasingly important as companies seek to turn AI investment into lasting business value without exposing themselves to unnecessary risk.