AIM Intelligence's Seoul AI Risk Conference Flags Agent Data-Command Conflation Risk

AIM Intelligence, a Seoul-based AI security startup, hosted the "AI Risk Conference Seoul 2026" on September 7 at the Samsung Financial Campus, drawing roughly 100 industry attendees from Korea and abroad. Seoul National University professor Lee Byung-young told the conference that AI agents fold attacker-planted data into the same LLM context as legitimate instructions, arguing the practical fix is architecturally separating external data from user commands rather than filtering content after the fact. AIM Intelligence CEO Yoo Sang-yoon said agentic and physical AI systems carry risks the industry hasn't previously had to manage, and Financial Security Institute AI-security research head Kim Sung-woong said Korean financial firms are moving beyond customer-facing chatbots into AI agents that execute transactions directly, which raises the need for independent security review beyond standard audits. Speakers from Microsoft and OpenAI also presented, on proactive vulnerability discovery and layered technology-governance-security design respectively.

The framing matters because it applies a familiar prompt-injection problem to agentic execution rather than static chat output: agents now take real actions — transactions, tool calls — based on context that mixes trusted commands with unvalidated external data. As Korean financial institutions push agents toward direct transaction execution, the attack surface shifts from information disclosure to autonomous action, which existing compliance-style security review isn't built to catch. AIM Intelligence's own position is notable here — the startup already counts LG, Hyundai Motor, and Naver's D2SF among its investors and is an early partner in Microsoft's MDASH AI security system, meaning its investor base overlaps with the enterprise buyers most exposed to this exact failure mode.

Builders shipping agentic products, especially any touching financial transactions or physical-world actuation, should treat the instruction-versus-data boundary as an architectural requirement rather than a downstream filter, per the SNU framing offered at the conference. Investors tracking the AI-security tooling category should watch whether demand genuinely expands beyond text into audio and video, as AIM's CTO Park Ha-eon indicated on stage, since that would signal security spend following agents into multimodal deployment rather than staying confined to chat interfaces.