VIDRAFT's Chitos Engine Builds Full Coverage for Korea's 17-Item Financial AI Security Evaluation Standard
TL;DR: VIDRAFT's AI security scanning engine, Chitos, has built dedicated modules covering all 17 checklist items in the Korea Financial Security Institute (FSI)'s newly published "Financial AI Agent Security Evaluation Criteria" — the standard that becomes mandatory for financial-sector AI red-teaming in 2027. The engine runs fully on-premises inside air-gapped networks, achieving benchmark parity with its cloud-connected counterpart, which matters enormously for regulated environments. Developers building AI agents for Korean financial or public-sector deployments should start mapping their threat models against this framework now.
What it is
Chitos is VIDRAFT's AI-augmented security inspection engine. Its latest module set targets the FSI's "Financial Sector AI Agent Security Evaluation Criteria," a framework published in October 2026 that organises security requirements into 6 domains and 17 checklist items.
Key facts from the source:
- The FSI framework will run as a pilot verification program in 2026, then become the official AI red-teaming standard for Korean financial institutions starting 2027.
- Of the 17 items, 8 address novel threat vectors specific to the agent execution layer, including:
- Goal integrity assurance
- Cascading-failure prevention
- Tool privilege control
- Sandbox isolation
- Inter-agent communication security
- Human-in-the-loop approval workflows
- The remaining items cover more traditional AI/application security concerns mapped to the six-domain structure.
VIDRAFT has also released a companion service called FreeWeb Security SCAN (operated under the Quant ID brand), a lightweight, no-login web vulnerability scanner built on the Chitos engine.
How it works
Chitos approaches AI security validation in a layered architecture that combines a conventional static/code scanner with an AI verification layer on top.
The 17-item module is structured as:
| Inspection mode | Count |
|---|---|
| Fully automated checks | 10 |
| Partially automated checks | 6 |
| Manual / control checks | 1 |
A critical design constraint is read-only operation: the engine does not interact with or modify the target system, making it safe to run against production infrastructure inside internal networks. All inspection happens on-premises, within the customer's own network boundary.
After a scan completes, the engine generates a structured diagnostic report that includes:
- Threat scenario analysis per checklist item
- A three-tier remediation guide (prioritised improvement steps)
- Regulatory clause mapping (linking findings directly to the FSI evaluation criteria items)
- OWASP Benchmark results alongside the FSI-specific output
The AI verification layer's primary function is false-positive suppression: it re-evaluates raw findings from the underlying code scanner and filters out noise before results surface to the analyst. According to VIDRAFT, this is what drives the performance gap between running the base scanner alone versus running Chitos.
FreeWeb Security SCAN operates on a lighter, passive-inspection model — URL input only, no authentication, no attack traffic. It checks security headers, HTTPS configuration, anti-email-spoofing settings (SPF/DKIM/DMARC class checks), and access permissions for nine major AI web crawlers, and delivers results within approximately one minute.
Benchmarks & results
All numbers below come directly from the source article (VIDRAFT's own self-reported measurements on the public OWASP Benchmark, scored via the Youden Index, 100-point scale):
| Configuration | OWASP Benchmark Score |
|---|---|
| Base code scanner alone | 52.8 / 100 |
| Chitos with external API engine | 77.2 / 100 |
| Chitos with on-premises self-hosted LLM engine | 77.8 / 100 |
Key takeaway for engineers: the performance gap between the cloud-connected and fully air-gapped on-prem deployments is effectively negligible (~0.6 points). This is the headline engineering claim — that network-isolation compliance does not cost you diagnostic accuracy. The AI verification layer is credited with blocking approximately 60% of false positives from the base scanner, which accounts for the ~25-point jump from the baseline.
Note: These are VIDRAFT's self-reported benchmark measurements. Independent third-party reproduction has not been cited in the source.
How to try it
FreeWeb Security SCAN is the only publicly accessible surface mentioned in the source. It is described as a publicly released web service — no login or account required, operated under the Quant ID brand. The source does not provide a direct URL for the service.
The Chitos on-premises engine and the 17-item FSI compliance module are enterprise/institutional offerings. No public Hugging Face repository, GitHub link, pip package, or OpenAI-compatible API endpoint is mentioned in the source article. Access would require direct engagement with VIDRAFT's enterprise sales channel.
FAQ
Q: Is Chitos relevant outside Korea, or is it specific to Korean financial regulation?
A: The FSI compliance module is purpose-built for Korea's financial regulatory framework (FSI 2027 AI red-teaming standard). However, the underlying OWASP Benchmark scoring and the agent-layer threat categories (sandbox isolation, tool privilege control, inter-agent communication security) map directly onto internationally recognised AI security concerns — making the conceptual framework broadly applicable even if the regulatory mapping is Korea-specific.
Q: Does the on-premises engine require sending any data to external endpoints?
A: According to VIDRAFT, the on-premises (air-gapped) engine operates entirely within the customer's internal network. The source explicitly states that financial and public-sector institutions under network-separation regulations can run scans "without any data leaving the internal network," and that this configuration achieves benchmark parity with the external API-connected version.
Q: When does the 2027 FSI AI red-teaming standard actually take effect, and what is the 2026 pilot?
A: The FSI launched a pilot verification program in 2026 to allow institutions to test against the 17-item criteria ahead of mandatory adoption. Full enforcement as the official AI red-teaming standard for the Korean financial sector is scheduled for 2027.
Originally reported by 파이낸셜뉴스 (2026-10-30) — source article.
Top comments (0)