
Corma raises $125M to train cybersecurity-optimized foundation models - SiliconANGLE
Corma.ai Ltd., a startup developing AI models for cybersecurity teams, has raised $125 million in a Series A round led by NEA, valuing the company at $1 billion.
The company, with offices in San Francisco and Tel Aviv, has identified a critical gap in current AI capabilities for cybersecurity. Corma recently tested several large language models by asking them to hack a simulated corporate network. The models succeeded in installing malicious files across 85% of test runs but only found 12% of the malicious files they themselves had installed.
Performance Advantage
According to Corma, this imbalance between offensive and defensive capabilities gives hackers an advantage over cybersecurity teams. The company's algorithms can analyze data from various cybersecurity tools including firewalls and network monitoring software, not only identifying threats but also generating remediation guidance for each one.
Since we came out of stealth, demand from the world's largest organizations has been overwhelming, and one of our hardest problems has been securing capacity fast enough to keep up with the demand.
Autonomous Operations
Corma's models can perform breach investigations autonomously, alerting human team members only when necessary. The system can also be manually controlled through a Slack chatbot, allowing administrators to request specific actions like verifying successful malware removal from compromised servers.
The company benchmarked its models against competitors using DefenseBench V5, which measures reliability and speed in defensive cybersecurity tasks. Corma's model outperformed GPT-6 Astra, Claude Opus 5.5 and several other cutting-edge LLMs in these tests.