Chitos: VIDRAFT's Autonomous Security AI That Moves from Static Scan to Live Exploit Proof
TL;DR: Chitos is a free, no-signup autonomous security AI from Korean AI startup VIDRAFT, powered by their Darwin-398B-JGOS model. Unlike conventional scanners that hand you a list of suspected vulnerabilities, Chitos runs a three-phase pipeline — static analysis, agentic research, and active exploitation — to deliver only verified, evidence-backed findings. If you do penetration testing or security auditing, this changes what "automated scanning" means.
What it is
Chitos is an autonomous security AI released by VIDRAFT (also known for the FINAL-Bench research group) on June 29, 2026. It is designed to close the gap between detection and proof — the step that traditionally requires a human penetration tester to confirm whether a flagged vulnerability is actually exploitable.
Key characteristics:
- Engine: Darwin-398B-JGOS, VIDRAFT's 398-billion-parameter model
- Access: Fully free, no account or sign-up required
- Input modes: Paste source code directly, or provide a URL
- Language support (Phase 1): Python, JavaScript/TypeScript, Go, Java, C/C++, PHP, Rust, YAML, and more
- Processing: Code is processed on VIDRAFT's own infrastructure — it is not forwarded to third-party AI services
The core design philosophy: unverified vulnerability candidates are treated as hypotheses, not findings. Only candidates that survive active validation appear in the final report.
How it works
Chitos operates as a three-phase pipeline:
Phase 1 — Static Security Scan
The system applies 50 language-agnostic pattern rules to detect a broad set of vulnerability classes: injection sinks, deserialization gadgets, credential leakage, path traversal, cryptographic weaknesses, CORS misconfigurations, prototype pollution, and Kubernetes privilege escalation, among others. Each candidate is cross-referenced against live CVE data. Darwin-398B-JGOS then assigns a CWE classification and a 0–100 security score.
Phase 2 — Autonomous Research (Web-Search Agent)
For questions that static code analysis cannot answer — e.g., "Does a known exploit chain exist for this library version?" or "Has this authentication pattern been bypassed in real-world incidents?" — Chitos spins up a web-search agent. It queries CVE databases, exploit repositories, and PoC listings autonomously. This phase requires a user-supplied Claude API key, which is used only for the duration of the request and is never stored.
Phase 3 — Active Exploit (Proof)
This is where Chitos diverges most sharply from conventional scanners. Against targets the user owns or has explicit authorization to test, Chitos sends real attack payloads. The pipeline covers:
- XSS: reflected, stored, and DOM-based
- SQL injection: time-based blind, error-based, UNION extraction
- SSRF, path traversal, authentication bypass, IDOR
The automation loop is: payload generation → response observation → mutation and retry on block → evidence capture → kill-chain construction → remediation priority ranking.
Verification layer
To suppress false positives — a chronic problem in security tooling — Chitos applies a mandatory re-verification pass before reporting. A candidate must satisfy three checks: Is it reachable from an actually exploitable code path? Is the response delta statistically significant? Did the payload complete a round trip? Candidates that fail any check are dropped.
This verification design draws on VIDRAFT's AETHER meta-cognitive adapter research, a system built to predict its own errors before they occur. That same "self-doubt" capability is repurposed here to filter noise before it reaches the report.
Benchmarks & results
The source article does not publish aggregate benchmark tables for Chitos itself. The one quantitative data point cited is from the underlying AETHER meta-cognitive research: the system achieves 90.3% accuracy at predicting its own errors on the DOUBT benchmark. This self-error-prediction capability is what underlies Chitos's false-positive suppression architecture.
No exploit success rates, scan speed figures, or comparative scanner benchmarks are publicly available at the time of writing.
How to try it
Chitos is publicly accessible right now with zero friction:
- Navigate to https://chitos.vidraft.net
- Paste your source code or enter a target URL
- Phase 1 (static scan) runs immediately — no credentials needed
- For Phase 2 (autonomous research), supply your own Claude API key when prompted (used in-session only, never stored)
- For Phase 3 (active exploitation), you must have explicit authorization for the target — this is your legal responsibility
No Hugging Face model card, GitHub repository, or OpenAI-compatible API endpoint for Chitos has been announced publicly at this time.
FAQ
Q: Is Phase 3 (active exploitation) legal to run against any target?
A: No. Sending real attack payloads to systems you do not own or have written permission to test is illegal in most jurisdictions. Chitos is designed for authorized penetration testing and bug bounty scenarios. The responsibility for authorization rests entirely with the user.
Q: Why does Phase 2 need my Claude API key — can't Chitos use its own?
A: The web-search agent in Phase 2 invokes Claude for reasoning over retrieved threat intelligence. VIDRAFT uses your key so that the request is billed to your account and your code is not relayed through a third-party service under VIDRAFT's credentials. The key is held only for the lifetime of the request.
Q: Does Chitos send my code to external AI providers?
A: According to VIDRAFT, code is processed on their own infrastructure using Darwin-398B-JGOS and is not forwarded to third-party AI services. The Claude API key used in Phase 2 is the sole external call, and it operates on metadata/research queries rather than your raw source code.
Originally reported by AI Chronicle (日本) (2026-07-01) — source article.
Top comments (0)