Gray Swan AI

Gray Swan AI builds tools for finding and closing security gaps in large language models and AI agents before attackers do. Its core products are Shade, an automated red-teaming agent that runs adversarial tests against models throughout pre-deployment and CI/CD, and Arena, a continuous public competition that draws more than 15,000 researchers and security professionals to attack AI systems and surface novel jailbreaks and exploits. The company also offers Cygnal, a runtime protection layer.

The company was spun out of Carnegie Mellon University by Matt Fredrikson (CEO) and Zico Kolter (chief scientist), both machine-learning security researchers, and is based in Pittsburgh. Its key differentiator is treating adversarial testing as a continuously-updated, crowd-sourced discipline rather than a one-time audit: new attack techniques found in Arena feed directly into Shade’s test suite. That approach has earned it citations in system cards from Anthropic, OpenAI, and Meta — a notable form of independent validation, since frontier labs do not credit external red-teamers lightly.

In mid-2026 the company closed a $40 million Series A co-led by Wing Venture Capital and Madrona, with Snowflake Ventures, Samsung Next, and Hudson River Trading also participating, to expand lab partnerships and go-to-market operations.