Put AI Agents into Production without Losing Control with Fortinet

As AI reaches production, agents act independently. FortiAIGate, a self-hosted AI gateway, stops prompt injection and data leaks. Roll out AI faster, with policy, spend, and data under your control. Patrick Bedwell of Fortinet’s product marketing organization presented FortiAIGate at Security Field Day 16. He described it as a single control point between AI applications and the models they call.

Bedwell opened with survey data. Nine in ten organizations have adopted AI, 20% have suffered a model-related data breach this year, roughly half have overspent their AI budgets, and three-quarters worry about data sovereignty. Traditional tools inspect payloads and files, so they miss attacks written in plain language, such as jailbreaks, poisoned tool calls, and denial-of-wallet prompts. FortiAIGate works as a secure reverse proxy and runs on Kubernetes clusters with NVIDIA GPUs. It inspects every prompt, response, and MCP tool call. Its guardrails include technology from a company Fortinet acquired this summer, and they detect sensitive data across 35 entity types, nine languages, and five media types. Customers can test and modify detectors before deploying them. They can also start from more than 50 governance templates, including HIPAA and GDPR, or import a policy document to generate guardrails. Violations can be blocked, redacted, replaced with realistic fake data, or flagged with an alert. Workspaces enforce access rules and hard budget caps per user, project, and model, and those budgets can reset daily or monthly.

Bedwell’s clearest example was an email-summarizing agent hijacked by an indirect prompt injection. Guardrails that check only the first prompt and the final response would miss that attack, but FortiAIGate sees every step in between. Full logging also gives auditors proof of what an application did with customer data. Bedwell was candid about the current limits. FortiAIGate does not see shadow AI traffic to SaaS tools, which FortiGate covers instead. Substituted data cannot be swapped back in the model’s response. The product requires NVIDIA GPUs. Performance figures for the newly integrated guardrails will be published after testing. He placed FortiAIGate within Fortinet’s broader AI protection strategy: FortiGate at the perimeter, FortiCNAPP for workloads and pipelines, FortiWeb and FortiAppSec for APIs, and FortiAIGate guarding the model itself.