Turning adversarial research into security tools
Gray Swan AI builds testing and runtime-security products for enterprises and model developers. Its current offering combines automated adversarial evaluation with a community of people testing model behavior. The company's researchers have academic backgrounds, but those affiliations do not make the business a university lab or a university-owned organization. [1]
Cygnet and the public launch
At its July 2024 public launch, Gray Swan made Cygnet available through a free web interface and a paid API. It described Cygnet as a Llama 3-based model tuned to resist attacks, alongside Shade, an evaluation and continuous-integration product. The announcement did not release downloadable Cygnet weights. [2]
Testing and protection are distinct
Adversarial evaluation tries to expose failures before or during deployment; runtime protection attempts to constrain behavior while an application is operating. Gray Swan sells capabilities in both areas. Its research and demonstrations explain the approach, while resistance claims still depend on the attacks, settings and models evaluated. [1][2]
Sources
- About Gray Swan
Gray Swan AI · retrieved Oct 23, 2026
- Public Launch
Gray Swan AI · Jul 16, 2024