South Korea's financial sector is gearing up for mandatory AI red-teaming in 2027 — and one startup is already compliance-ready.
TL;DR: VIDRAFT, a Korean AI security startup, has built dedicated compliance modules inside its Chitos security inspection engine that address all 17 evaluation criteria published by the Korea Financial Security Institute (FSI) for AI agent security. The FSI framework, organized across six domains and 17 checkpoints, is scheduled to become the official standard for financial-sector AI red-teaming from 2027. VIDRAFT simultaneously unveiled a free web vulnerability scanning service, FreeWeb Security SCAN, powered by the same Chitos engine.
VIDRAFT (CEO Kim Min-sik), a Korean AI security startup, announced on October 29, 2026 that its proprietary security inspection engine, Chitos, has been fully mapped to all 17 evaluation criteria outlined in the Korea Financial Security Institute's newly released "Financial Sector AI Agent Security Evaluation Standard." The announcement, first reported by 파이낸셜뉴스, signals VIDRAFT's intent to lead compliance readiness as South Korea's financial industry prepares for a significant regulatory shift in AI oversight.
The Korea Financial Security Institute released its AI agent security evaluation guidelines this month. The framework spans six domains and 17 individual checkpoints. It is currently in a pilot verification phase and is set to become the official benchmark for AI red-teaming — structured security testing that simulates adversarial attacks — across the Korean financial sector starting in 2027.
Of those 17 checkpoints, eight are classified as newly identified threat vectors specific to AI agent execution layers. These include verifying goal integrity, preventing cascading failures, controlling tool-level permissions, enforcing sandbox isolation, securing inter-agent communications, and validating human-approval workflows. The breadth of these criteria reflects how much more complex the attack surface becomes when financial institutions deploy autonomous AI agents rather than static models.
To address all 17 criteria, VIDRAFT built a dedicated module set within Chitos. The breakdown covers ten fully automated checks, six partially automated checks, and one control-based inspection. The entire process runs in a read-only mode directly on a client's internal network, meaning the engine does not alter, interrupt, or otherwise affect live systems during testing. Once a scan is complete, the engine generates a diagnostic report that includes threat scenario analysis, a three-stage remediation guide, and mappings to specific regulatory provisions.
VIDRAFT also noted that Chitos produces results aligned with the internationally recognized OWASP Benchmark, giving financial institutions an additional reference point that goes beyond domestic regulatory requirements.
On the performance side, VIDRAFT reported that Chitos achieved a score of 77.8 out of 100 on its own OWASP Benchmark measurement (using the Youden index). The company explained that a standalone code scanner without the AI verification layer previously scored 52.8 points, and that the addition of an AI-based filtering step reduced false positives by 60%, lifting the score into the 77-point range. Notably, the company stated that its cloud-connected engine and its fully on-premises, air-gapped engine achieved virtually equivalent scores — 77.2 and 77.8 respectively — making the solution equally viable in network-isolated environments typical of Korean financial and public-sector institutions.
In a separate but related announcement, VIDRAFT launched FreeWeb Security SCAN, a no-login, no-attack web vulnerability diagnostic service built on the Chitos engine. Users enter a website URL and receive results within one minute covering security headers, HTTPS configuration, email spoofing prevention settings, and whether access is permitted to nine major AI web crawlers.
A VIDRAFT spokesperson said demand is growing for inspection solutions that operate stably inside institutional internal networks ahead of the 2027 rollout of formal AI red-teaming standards, and that the company intends to use its on-premises, network-isolated engine as its primary tool to address AI security verification needs across the financial and public sectors.
South Korea's financial regulators are moving deliberately toward institutionalizing AI security testing, and the 2027 deadline for mandatory AI red-teaming compliance is fast approaching. For financial institutions that operate under strict network-separation rules — preventing internal systems from connecting to external services — finding a viable AI security inspection tool has been a genuine challenge. VIDRAFT's claim that its on-premises engine performs at parity with its cloud-connected counterpart directly targets that gap.
The fact that eight of the 17 FSI criteria are newly defined threat categories specific to agentic AI systems also reflects a broader industry recognition: as financial firms deploy AI that can take autonomous actions, traditional security frameworks are no longer sufficient. Tools built from the ground up with these agentic threat models in mind represent a meaningful step forward for the sector.
Q: What is the Korea Financial Security Institute's AI agent security evaluation standard?
A: It is a framework of 17 checkpoints across six security domains, published in October 2026, that will serve as the official standard for AI red-teaming in South Korea's financial sector beginning in 2027.
Q: How does VIDRAFT's Chitos engine handle network-isolated financial environments?
A: Chitos offers a fully on-premises installation option that runs directly on a client's internal network in read-only mode, allowing institutions operating under network-separation regulations to conduct AI security inspections without transmitting data externally.
Q: What is FreeWeb Security SCAN?
A: It is a free web vulnerability diagnostic service powered by VIDRAFT's Chitos engine that delivers results within one minute by scanning a submitted URL for security headers, HTTPS status, email spoofing protections, and AI crawler access settings — requiring no login and performing no attacks on the target site.
Source: 파이낸셜뉴스 (2026-10-30) — original article