Same Model for Offense and Defense, Corma Raises $60 Million to Train AI Defense

According to NetSense Beating monitoring, cybersecurity startup Corma has officially debuted, completing a $60 million seed round with Sequoia leading the investment. The company is training a custom-built base model for network defense, which is then developed into a security Agent that can directly interface with enterprise's existing security tools.

Corma also conducted a series of red team/blue team tests. In a simulated enterprise network, Claude Opus 4.8, GPT-5.5, Grok 4.3, and DeepSeek V4 took turns attacking and defending, totaling 241 rounds. The attackers successfully implanted persistent backdoors in 85% of the tests, which remained undetected even after system reboots, while the defenders only discovered 19% on average. Interestingly, when the models were tasked to find the backdoors they themselves planted, 78% went undetected.

Corma believes that while attacks mainly test code and reasoning, defense is more akin to finding a needle in a haystack, requiring long-term anomaly detection from logs, traffic, and regular activities. This particular capability is not the primary focus of general model training.

However, Corma has not yet disclosed its model's performance in the same set of tests. The company claims that its proprietary model far surpasses general models, but a direct comparison is not yet visible.