Visa Vulnerability Agentic Harness for Project Glasswing (github.com)

Visa has announced the launch of the Visa Vulnerability Agentic Harness (VVAH), an open-source framework designed for autonomous vulnerability discovery leveraging advanced AI models. Developed in conjunction with learnings from Project Glasswing, a research initiative by Anthropic, VVAH improves the process of identifying and triaging software vulnerabilities by emphasizing pre-analysis threat modeling, multi-agent deterministic voting to minimize false positives, and structured artifacts to enhance the transition from discovery to action. Notably, the main focus of VVAH is on reducing the Mean Time to Adapt (MTTA)—the time between detecting a weakness and implementing a fix—addressing the critical bottleneck in AI-assisted vulnerability management. Technically, VVAH is compatible with multiple AI models, including Anthropic Claude and OpenAI’s tools, allowing users to choose the best-fit solution without dependency on any single provider. The framework operates in three phases—Discovery & Modeling, Deep Dive & Verification, and Synthesis, Chaining & Reporting—divided into nine stages that ensure a systematic approach to vulnerability analysis. Each stage incorporates reusable skills that can be modified independently, promoting efficiency and adaptation in deployment. With this release, Visa positions itself at the forefront of AI-enhanced cybersecurity solutions, promising heightened accuracy in vulnerability management while urging responsible use of the tool.