As enterprise AI deployments become more widespread, security is naturally taking center stage in the conversation. Many discussions focus on securing the model itself, but this overlooks an important point — that AI is changing the way traffic flows across the network; and this has security implications that need addressing too.
Data in motion is vulnerable, particularly when it’s crossing boundaries between systems and networks. Since AI deployments are increasingly distributed, data is moving constantly between devices, locations, clouds, and data centers.
Protecting moving data isn’t a new challenge. Enterprises have been wrestling with the need to secure data moving through porous network perimeters ever since cloud and SaaS applications came into common use. But AI-related data flows behave differently, and this is beginning to throw up some novel challenges.
AI Means More Data and New Flow Shapes
The first issue is the sheer quantity of data. Cisco predicts that the adoption of agentic AI could grow network traffic by about nine times by 2035, driven by autonomous task execution and inference-heavy workflows. More data means more risk and more serious consequences if there’s a breach.
The second point is that data moved by AI workflows can take labyrinthine paths across diverse and fragmented networks. These complex journeys are hard to secure end-to-end and open up more opportunities for attackers to exploit.
Finally, the shape and direction of data flows is beginning to change.
AI inference flows are much longer-lived than traditional non-AI web traffic, lasting twice as long. These need to be tracked and secured from beginning to end — a task that even next-generation security systems can struggle with.
East-west traffic is increasing too. Models are beginning to move closer to the users and devices that generate the data AI ingests and receive the output it generates. Robots in a factory, for example, might send data to a local AI model rather than a centralized one to keep response times low. The local model analyzes the information and sends instructions back to optimize robot performance and avoid bottlenecks in the production line.
Traditional perimeter-based security models generally don’t inspect east-west traffic. Without effective monitoring of internal traffic, though, attackers can quickly move through systems and exfiltrate data without detection, thanks to blind spots.
Stateful and Session-Based Security Systems Are Essential…
Traditional security controls have difficulty telling the difference between legitimate AI traffic and harmful activity like data exfiltration. That’s why many enterprises rely on stateful and session-based security, like next-generation firewalls (NGFW) and intrusion prevention systems (IPS) to track network flows.
Instead of monitoring packets in isolation, stateful security systems track entire active flows and use contextual information to work out whether they contain anything that’s potentially harmful. This helps to identify threats like multi-turn attacks, which are made up of small steps that appear harmless if looked at individually but gradually escalate into something more malign.
… but Longer-Lived Sessions Threaten the Performance of Security Appliances
But there’s a snag with this approach. Tracking active network sessions uses more CPU and memory than stateless security systems do. And since inference sessions last longer, there are likely to be more concurrent connections to be tracked, maintained, and inspected compared to non-AI flows.
That puts pressure on security appliances. If these become overwhelmed, latency could increase or security teams might be forced to maintain performance by reducing inspection levels.
Security Systems Must Become More Scalable
Many firewalls have threat defenses built into them, like IPS and data loss prevention (DLP), avoiding the need for separate systems. This helps detect and mitigate threats quickly and improves security posture, but it can come at a cost to performance.
Many firewalls operate by working through a sequence of security functions, which slows throughput. NGFWs that use parallel processing — where each packet is inspected once for multiple security functions — overcome this problem and allow more security services to be added to the firewall without crippling performance.
Cloud-delivered security, like secure access service edge (SASE) and Firewall-as-a-Service (FWaaS), may offer more flexibility and scalability compared to non-cloud options. The elasticity of cloud takes the burden of monitoring long-lived inference flows off local edge routers and allows extra capacity to be dialed up when AI traffic is putting a lot of pressure on the network.
SASE is designed to eliminate blind spots and bring identity-based security controls to the point of access, helping secure today’s porous enterprise perimeters. Microsegmentation limits lateral movement if attackers gain entry, reducing the fallout from breaches.
Securing East-West Traffic Shouldn’t Compromise Latency
SASE might be quickly becoming the gold security standard for disparate WANs, but the picture can be a little more complex when it comes to protecting east-west traffic.
If east-west traffic has to be routed to a local SASE point of presence (PoP) for inspection and then returned to the local network, the extra latency can impact how well AI models function.
The compromise is to deploy SASE to secure data flowing across the WAN and use NGFWs to secure east-west traffic. But if these are deployed and operated as separate systems, security is fragmented, and managing policies, threat detection, and telemetry is much more complex and time-consuming. That’s why it’s important to choose firewalls and SASE offerings that run on the same operating system and can be managed as a coherent whole.
Tackling Infrastructure Complexity Is the First Step to Scaling AI
The reality is that most enterprise network and security architectures are complex and fragmented with poor visibility. Combined with the fact that AI requires security and governance controls operating across multiple architectural layers, this complexity makes securing growing numbers of AI data flows a daunting prospect.
A good place to start is unifying security and disparate network connections from multiple carriers into a single managed platform to simplify infrastructure and plug the gaps that attackers can exploit. This improves visibility and accountability and reduces the burden of management, helping IT teams scale AI deployments without compromising performance or protection.