Multi-model AI security strategies: Two steps to take now

Lee EnnisSr Manager, Data Science | F5

AI security is dominating industry conversations this year, but charting the safest path forward can feel overwhelming. The rapid release of ever-more powerful models, the surge in agentic traffic, the rise of shadow AI, and the threat of frontier models in adversarial hands loom heavy over both daily operations and long-term roadmaps.

Leaders must also think strategically about resource allocation when deploying multiple AI models. Just as you wouldn’t drive a Porsche 911 Turbo for mundane errands, you shouldn't rely on your most expensive, compute-heavy models for simple tasks. Maximizing AI efficiency is critical, yet stepping back from the daily firehose to gain that broader perspective remains a challenge.

That isn’t slowing AI adoption, of course. Our 12th annual survey of global IT decision-makers reveals just how deeply AI apps are already embedded in production systems and operational workflows. In enterprises with at least $1 billion in annual revenue, an average of 55% of apps are already AI-enabled, surpassing the halfway mark in just a few short years. Meanwhile, 78% of organizations are managing their own inferencing, at least in part.

“Ready or not” isn’t safe

Unfortunately, the data also suggests that many organizations are getting ahead of themselves. There is little question that AI adoption has outrun visibility. More than one-third of our survey respondents (35%) admit that their infrastructures aren’t ready to deliver the governance, observability, and security their AI workloads need. Others are likely speeding into the fog with a false sense of control and no appropriate headlights.

Luckily, there are several key actions you can take today to ensure your organization proceeds safely toward AI success.

Step 1: Grab the wheel of multi-model AI

The first step is to take control of your multi-model deployments. The average organization uses seven different models or model families in production or active evaluation, and the largest enterprises, including many F5 customers, typically use even more. While 52% of organizations are chaining or orchestrating models, nearly half are operating multiple models with little coordination and inconsistent security.

Before AI models proliferate further, inventory your models and inference services. Treat them as a new infrastructure layer that must be secured and optimized:

• Identify redundancies: Spot overlapping capabilities and resolve compatibility issues early.
• Analyze cost drivers: Understand exactly where your compute budget is going.
• Assess security risks: Pinpoint vulnerabilities specific to certain models or workloads.
• Align tasks to models: Ensure your AI models are performing the tasks that will help your business thrive, while your more mundane needs are met by more workaday services at costs that align with their value.

Tools like model registries can help clarify dependencies and enable orchestration frameworks for seamless model switching or failover. A comprehensive inventory builds efficient routing and resilience, delivers consistent security, and makes it easier to scale without cost, complexity, or security pitfalls.

Step 2: Navigate the AI traffic surge

Your inventory will also prepare you for the next major hurdle: the explosion of AI agents. Automated and AI-driven traffic already represent more than half of Internet traffic. With two-thirds of organizations using AI to drive further automation, the number of agents on the web is growing exponentially. In fact, 77% of organizations already expect agent identity and access control to create operational and security challenges.

They are right to worry. Managing access for machine identities can quickly swamp security administrators, bringing a significant risk of credential theft, privilege abuse, and impersonation.

Prevent disasters by developing strict governance policies for non-human identities. Effective approaches include routing traffic via identity-aware load balancers and implementing guardrails that limit autonomous AI actions to reversible performance optimizations. Securing non-deterministic actions can be challenging, but platform security solutions like our recently announced F5 AI Security Platform eliminate the guesswork.

When establishing governance, be sure to include economic modeling for agent service consumption. Revenue models are still in flux, and per-task or per-interaction charges may soon replace traditional per-user fees. Just as seatbelts and rollbars add security for off-road adventures, addressing these governance issues now will prepare your organization for wild agent traffic without disrupting future workflows.

These are just a few steps organizations should consider to improve multi-model AI visibility and cybersecurity amid explosive agentic growth. To learn more, read our recent report: “The State of the AI Inferencing Landscape in 2026: Implications for App Infrastructures.”

About the Author

Lee EnnisSr Manager, Data Science | F5

Lee Ennis leads the AI data science and research team at F5, where he focuses on developing advanced AI security solutions that help enterprises safely deploy and monitor large language models (LLMs) and generative AI systems. His team combines expertise in machine learning, adversarial testing, and model assurance to identify vulnerabilities and strengthen trust in AI-driven environments.

More blogs by Lee Ennis