**[The Same Model Excels at Attacking but Struggles at Defending, Corma Secures $60 Million to Train Defense AI]** According to monitoring by Beating, cybersecurity startup Corma has officially launched and completed a $60 million seed funding round, led by Sequoia. The company is retraining foundational models specifically for cybersecurity defense and transforming them into security agents that directly integrate with existing enterprise security tools. Corma also conducted a series of attack-defense tests. Claude Opus 4.8, GPT-5.5, Grok 4.3, and DeepSeek V4 took turns attacking and defending within the same simulated enterprise network, running a total of 241 tests. Attackers successfully left persistent backdoors after system reboots in 85% of the tests, while defenders only detected an average of 19%. Even when tasked with finding the backdoors they themselves planted, the models failed 78% of the time. Corma believes that attacking primarily tests coding and reasoning skills, whereas defending is more akin to finding a needle in a haystack—requiring long-term analysis of logs, traffic, and normal activities to identify anomalies. These capabilities are not the focus of training for general-purpose models. However, Corma has yet to disclose its own model's performance in the same set of tests. The company claims its model has already surpassed general-purpose models, but direct comparisons are not currently available. [Original Link]