Chitos Covers All 17 Items of Korea's Financial AI Agent Security Criteria, Scores 77.8 on OWASP Benchmark On-Prem
securityaidevsecopsfintech
TL;DR
Chitos, VIDRAFT's AI security engine, now has a module for every one of the 17 items in the Financial Security Institute's new "AI Agent Security Evaluation Criteria for the Financial Sector" (Korea, October 2026), which become the official basis for AI red-teaming in Korean finance in 2027. On the public OWASP Benchmark, Chitos scores 77.8 (Youden index) with an engine that runs entirely inside the customer's network, against 52.8 for the base scanner alone. A free website check built on the Chitos engine, FreeWeb Security SCAN, is open to everyone.
Why 17 items matter
The criteria have 6 areas and 17 items. Eight are new threats that live in the agent execution layer, not in the model: goal integrity, cascading failure, unsafe tool use, tool permission control, sandboxing, agent-to-agent communication, human approval and human-in-the-loop overload. A model-level safety test does not cover them.
| Coverage | Items |
|---|---|
| Automated | 10 |
| Partly automated | 6 |
| Control review | 1 |
| Total | 17 of 17 |
All checks are read-only and non-destructive, run inside the customer's network, and end in a report with threat scenarios, three-step remediation and regulatory mapping.
OWASP Benchmark: 77.8, same on-prem as with an API
OWASP Benchmark has about 2,700 test cases and scores security tools on a 0 to 100 Youden index.
| Configuration | Youden |
|---|---|
| Base code scanner alone | 52.8 |
| Chitos, external API judge | 77.2 |
| Chitos, on-premises judge | 77.8 |
The AI verification layer removes more than 60% of false positives without dropping true findings. The on-premises judge matches the API judge, so data never has to leave the network, which is what regulated finance and public institutions need. Published commercial tool scores sit in the mid-70s and above. These are our own measurements; cross-checking with the official scoring tool is in progress.
FreeWeb Security SCAN
Enter a domain and get a report in about a minute on security headers, HTTPS, email spoofing protection, SEO basics and whether 9 AI crawlers can reach the site. It is read-only: no attack, no login. The service runs on the Chitos engine and is operated by QuantID.
FAQ
What are the FSI criteria? The Financial Security Institute's AI agent security criteria: 6 areas, 17 items, piloted in 2026 and applied to financial-sector AI red-teaming from 2027.
Does Chitos need to send data outside? No. The on-premises judge scored 77.8, the same level as the API judge at 77.2.
Are the OWASP scores independent? They are our own measurements. Cross-checking with the official scorer is in progress.
Is the free scan safe to run? Yes. It only reads what the site already exposes publicly.