Securing Your Own AI Agent Runtime: Lessons From the Fortinet-Virtue AI Deal

Fortinet's acquisition of Virtue AI, an AI runtime protection and validation specialist, confirms AI-specific security tooling has matured into a real acquisition category, not just a startup niche. Most engineering teams don't have a dedicated tool like this yet — here's a practical, tool-agnostic checklist for the gap it fills.

What AI Runtime Security Actually Covers

Traditional application security assumes deterministic behavior — the same input produces the same output, so testing and monitoring can be built around known code paths. AI agents don't behave that way: the same prompt can produce different action sequences, and the security-relevant surface includes the model's own reasoning process, not just its network and data access.

A Practical Checklist Before You Buy a Dedicated Tool

• Log every action an agent takes, not just its final output — a decision trail matters for AI agents the way it doesn't for deterministic code, since the same prompt can legitimately produce different action sequences
• Define explicit action allowlists per agent (what it can actually do, not just what data it can access) — this is the control that most directly would have prevented the specific failure modes seen in recent incidents (OpenAI's Hugging Face breach, AISI's unsanctioned actions)
• Set a maximum autonomous session length or action count before requiring human check-in, calibrated to the agent's actual risk tier — unlimited autonomous operation is the condition that let recent incidents compound before detection
• Test your own agents against the OWASP LLM Top 10 categories specifically, not just general security review — excessive agency and insecure output handling are the categories most directly relevant to autonomous agent runtime risk