Gray Swan Shade and Cygnal AI Security

Gray Swan Security Inc. was founded by AI safety and security researchers from Carnegie Mellon University, led by co-founder and CEO Matt Fredrikson and co-founder and Chief Scientist Zico Kolter. It provides adversarial testing and runtime enforcement for AI models and agents, using techniques that frontier labs rely on for pre-release safety evaluations, and reports more than 20 customers including frontier labs and global enterprises.

Shade runs automated behavior assessments in which LLM-based adversarial agents attack models, agents and guardrails in the customer's own deployment context. Cygnal classifies and blocks adversarial inputs and unsafe agent outputs at runtime with continuous agentic monitoring. Arena, a red-teaming competition network of more than 15,000 researchers with over three million attack attempts, feeds new attack techniques into both products. Gray Swan also offers pre-release adversarial evaluations for model builders and expert red teaming services; its evaluations are cited in system cards of major models.

• Shade: automated adversarial assessment using LLM-based attacker agents in the customer's deployment context.
• Cygnal: runtime classification and blocking of adversarial inputs and unsafe agent outputs.
• Arena: red-teaming competition network of over 15,000 researchers supplying new attacks.
• Adversarial evaluation: pre-release benchmarks and private red teaming for model builders.
• Agent and MCP coverage: testing and monitoring for tool-connected agent workflows.
• Research hub: over 150 tools and datasets for AI security research.

Vietnamese companies building their own language models can commission pre-release adversarial evaluations. Banks and fintechs deploying customer-service agents can run Shade assessments and add Cygnal runtime blocking. Software firms building agent products for export can use test results as security evidence for foreign clients.

Enterprises test their agents with the same attack methods used to evaluate frontier models before release. Attacks found by a community of over 15,000 red teamers keep test coverage current as new jailbreaks appear. Cygnal adds runtime blocking through a simple API integration with Python, JavaScript or cURL. Results give security and compliance teams evidence for MCP and tool-connected agent workflows.

Frontier labs and regulators now treat independent adversarial evaluation as a standard pre-release step, and enterprises deploying agents are adopting the same practice. Crowd-sourced red teaming is emerging as a way to keep pace with new attack techniques, relevant for Southeast Asian firms that build on global foundation models.

Gray Swan is a Carnegie Mellon University spin-out. In June 2026 it raised a 40 million USD Series A co-led by Wing Venture Capital and Madrona, with Obvious Ventures, Snowflake Ventures, Hudson River Trading, Samsung Next and Magarac Venture Partners; Snowflake is a strategic partner. Its site shows Cyber Essentials and SOC 2 badges and a trust center.