Corma is building a defensive cybersecurity foundation model to counter AI-enabled exploits, zero-days, and scaled social engineering

Corma is training a foundation model focused on enterprise defense as agentic AI improves attackers’ ability to find vulnerabilities and automate social engineering. The company argues that general-purpose frontier models are poorly suited to defensive cybersecurity because critical signals live in logs, telemetry, and traces rather than typical text training data. In its red/blue team tests, defenders failed to find a planted backdoor 78% of the time, even when using the same model that created it.

The pitch reflects a growing belief that security AI will require domain-specific models trained on operational data, not just wrappers around general LLMs. If offensive AI continues to lower attack costs faster than defenses improve, enterprises may shift spending toward specialized autonomous detection, investigation, and response systems.