Corma Raises $60M from Sequoia to Build Defensive AI for Cybersecurity
Corma has secured $60 million in seed funding from Sequoia Capital, Khosla Ventures, and Coatue to develop a groundbreaking defensive cybersecurity foundation model.
The Tel Aviv and San Francisco-based startup is creating an AI agent designed specifically for defense, already deployed at Fortune 100 and 500 companies. In simulations, however, the same AI models that successfully attacked failed to defend effectively.
August 10, 2026 – 8:30 pm
Summary:
Corma tested leading AI models like GPT and Claude across simulated Fortune 500 scenarios with dozens of existing security tools. The attackers won 88% of the simulations while the defenders only caught 12%. This highlights a critical gap in current cybersecurity defenses as AI capabilities for attack grow more sophisticated.
Corma aims to fill this void by building a dedicated defensive model from scratch, contrasting with software solutions that rely on existing tools. Their AI agents operate across an organization’s existing security infrastructure, automating tasks end-to-end.
Key Points:
- Unique Approach: Instead of replacing existing tools or functioning as a product, Corma offers "virtual human resources" deployed as needed by each organization.
- Impressive Results: Early deployments at Fortune 100 and 500 companies saw Corma reduce threat response times by over 94% and expand security coverage 15x.
- Experienced Team: The company draws on expertise from frontier AI researchers at Google and DeepMind, combined with cybersecurity specialists from Israel’s Unit 8200.
- Addressing a Growing Concern: Corma’s focus is on purpose-built defense as AI capabilities for both attack and defense continue to evolve rapidly.
The Bigger Picture:
Corma’s approach challenges the assumption that general-purpose AI can effectively handle both offensive and defensive cybersecurity. They argue that specialized models are needed for the unique structural demands of each side, given the distinct reasoning and processing needs.