Fortinet Korea details AI gateway plan at AIS 2026

Fortinet Korea managing director Kim Su-young presented an integrated AI security strategy covering large language models and Agentic AI at the AIS 2026 conference in Seoul on November 3, 2026.

Kim described attack surfaces including indirect prompt injection through long contexts, multi-turn conversations and instructions hidden in external documents. In retrieval-augmented generation environments, he noted, poisoning of reference documents or knowledge databases can let AI use malicious content in business decisions. Tool descriptions and responses connected via the Model Context Protocol, he said, must also be inspected for tool poisoning and indirect prompt injection.

Gateway architecture and FortiAIGate

Kim presented a gateway architecture that authenticates and inspects prompts before model calls and inspects outputs and sensitive data after responses. Fortinet's FortiAIGate was introduced with inspection covering system instructions, conversation history, tool definitions, past tool calls and responses, and model-generated tool arguments. Detection combines DLP, regular expressions and natural-language semantic analysis, applied through input and output guardrails by model, department and workspace policy.

On agent security, Kim described inspecting MCP tools and skill content before execution and checking prompts and tool calls during execution. He stressed inventorying agent assets and their connections, and managing non-human identities such as service accounts and API keys using virtual keys with usage, permission, budget and expiry control.

Kim recommended AI red teaming during operation and re-verification whenever models, prompts or connected tools change. For domestic deployment, he said Korean-language understanding, customer GPU and Kubernetes environments, and security inspection latency must be verified. The integrated strategy combines network, web/API and cloud infrastructure protection with AI input, output and action controls.

AIS 2026 was hosted by Daily Secu with support from the Korea Internet & Security Agency and the Korea Information Security Industry Association, and drew about 700 security staff and officials to the Korea Science and Technology Center in Seoul.