
Openwing: Janus is an autonomous DevSecOps platform built to tackle the growing security backlog caused by fast-paced, AI-assisted software development. While code can be generated quickly, fixing vulnerabilities still relies on slow, manual review and patching. Unlike traditional scanners that just highlight problems, Openwing: Janus combines offensive and defensive AI agents in a continuous feedback loop. This setup lets it analyze code, find vulnerabilities, suggest possible exploit paths, and create patches for verification. Powered by the Qwen2.5-72B-Instruct model running on vLLM with AMD hardware, it takes advantage of the AMD Instinct MI300X GPU’s 192GB memory to serve a 72-billion parameter model efficiently. The Red Team agent scans repositories using a structured vulnerability database to detect major issues like SQL injection or SSRF. Then, the Blue Team agent rewrites the risky code to strengthen security without breaking the app’s functionality. The patched code goes through another verification pass, creating a closed attack-and-defense cycle instead of a one-off scan. To keep things dependable for developers, Openwing: Janus uses a fault-tolerant approach with a fallback chain that mixes local AMD-hosted inference with external model providers. Its React-based frontend streams live agent activity, so users can watch scanning, exploit reasoning, and patching in real time. This not only makes it a powerful security tool but also offers a transparent development experience, helping teams see exactly how AI agents make their security decisions. The aim with Openwing: Janus is to show that advanced, agent-powered security workflows can be both practical and transformative for modern development teams.
10 May 2026