ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsQuick take3 min readDisconfirmAI Editorial

Kuntek CEO: AI Agent Security Must Connect the Full Lifecycle to Reduce Response Time

South Korean convergence-security firm Kuntek CEO Bang Hyuk-jun argues that securing AI agents requires connecting asset identification (extending software bill-of-materials to an AI-BOM), runtime policy enforcement, and hardware-isolated data protection into one measurable lifecycle, rather than treating them as separate products.

2026-11-20 16:37 GMT+8

When enterprises deploy AI agents, the first question is not which threat to block but what actually exists in the environment. Bang Hyuk-jun, CEO of South Korean convergence-security company Kuntek, stated in a November 2026 interview with Byline Network that AI agent security must begin from the same principle as traditional security: asset visibility. He noted that if an organization cannot accurately identify the servers and software it already holds, managing AI agent risk becomes more difficult still—because knowing which systems are affected is a prerequisite for finding vulnerabilities and responding to them.

Kuntek extends this identification into three layers. First, a software bill-of-materials (SBOM) catalogs the programs, libraries, and versions inside each software package; when a vulnerability is disclosed in a component, the SBOM helps trace which systems are affected. Second, an operational bill-of-materials (OBOM) checks what is actually running in live environments, not just what was declared at build time. Third, an AI bill-of-materials (AI-BOM) records which models are in use, where they came from, what licenses they carry, and what tools and permissions agents have connected to. Bang called this raising the "resolution of visibility" across the AI lifecycle. Kuntek's supply-chain product AEZIZ, according to the company, has expanded its SBOM management into the AI-BOM domain, adding checks for model provenance, dangerous file formats, and agent-connected tool permissions.

Identification alone does not eliminate runtime risk. Bang pointed to indirect prompt injection—where malicious instructions are hidden in documents an agent reads—as a concrete failure mode. A hiring AI that interprets hidden text in a resume as a legitimate work command could produce distorted evaluations. He noted that even when an agent uses a normal account and API, it can still perform a dangerous action; existing authentication and network access controls alone make it difficult to judge whether the agent's behavior matches its business purpose. Kuntek's answer is a platform called ASH, which checks each agent action against security policy before execution: if an agent attempts to send customer data to an external system, the platform verifies the purpose and permissions of that tool call, then allows, blocks, or escalates to human approval. The company describes ASH as being in proof-of-concept and early commercialization.

Sign in / Sign up
ReadingResearchRadarInvestment framework

For data-sharing scenarios—where enterprises need to let external AI developers compute on sensitive data without handing over originals—Kuntek proposes NEXA, which uses hardware-based trusted execution environments (TEE) to run pre-agreed computations inside a protected enclave. The company says it built this on PrivyBridge, a technology developed based on research from a Seoul National University team. A hospital and an AI firm could, in principle, run joint analysis without the hospital releasing raw patient records. NEXA is also described as being in PoC and early commercialization, with discussions underway with public institutions.

Bang's ultimate argument is that these three layers—AEZIZ for identification, ASH for runtime control, NEXA for data protection—must be connected into one flow so that a newly disclosed vulnerability can be traced to affected systems, checked for active exploitation, and remediated without manual handoffs between separate products. He proposed three time metrics for the whole chain: mean time to visibility (MTTV), mean time to detect threats (MTTD), and mean time to respond (MTTR). Repetitive judgment steps, he said, should be automated with AI—but not all of them. He distinguished human-in-the-loop, human-on-the-loop, and human-out-of-the-loop modes, with the boundary set by each organization based on task criticality and risk.

Every capability described is Kuntek's own account of its products and methodology. No independent evaluation, deployment benchmark, or third-party reproduction of the claimed effects is available in the source material. The conceptual framework—visibility first, then runtime enforcement, then connected lifecycle metrics—is a reasonable security architecture, but its effectiveness at scale remains the vendor's assertion, not a verified result.

Sources

  • byline.network바이라인네트워크↗
Continue researching →