Beating AI News Flash: Cybersecurity AI startup Corma has announced the completion of a $125 million Series A funding round at a valuation of $1 billion, led by NEA, with Sequoia Capital, Khosla Ventures, and Coatue participating.
The company was only founded last year and just emerged from stealth in August with a $60 million seed round. In just three months, its cumulative funding has reached $185 million.
Corma believes that existing large models are becoming increasingly adept at programming and at finding and exploiting software vulnerabilities, but defensive capabilities have not kept pace.
In August, Corma had GPT-5.5, Claude Opus 4.8, Grok 4.3, and DeepSeek V4 attack and defend against each other in a simulated enterprise network. In 241 experiments, the AI attackers had an 85% chance of successfully implanting persistent backdoors, while defenders only found the backdoors in 19% of experiments. Even when the same model was used for both attack and defense, defenders frequently failed to find what the attackers had left behind.
Corma's explanation is that attacking relies primarily on coding capability, while defending requires finding anomalies in massive logs, network traffic, and security alerts. General-purpose large models are not specifically trained on this type of data.
Therefore, Corma chose to train a defense-specialized foundation model from scratch and have the model repeatedly perform attack and defense tasks in simulated enterprise environments, improving its capabilities through reinforcement learning.
Corma has also turned the model into an AI security employee that directly uses enterprises' existing security tools to automatically investigate anomalies, track hacker activity, and even stop attacks in progress. According to the company, some Fortune 500 enterprises have already begun using the system. Over the past three months, Corma has expanded its server capacity 24-fold to meet growing demand.
