Corma Raises $60M for Defensive Cybersecurity AI

SiliconANGLE and The Next Web report that Corma ran hundreds of simulations modeled on Fortune 500-style enterprise environments containing dozens of security tools. The company first tasked leading models from OpenAI and Anthropic with attacking the simulated networks and planting persistent threats, then asked the same models to identify and remove those threats.

According to Corma's reported simulation results, AI attackers succeeded in 88% of cases, while defenders detected threats in 12%. These are company-run tests rather than an independently published benchmark, so the cited reports do not establish how the results generalize across models, enterprise telemetry stacks, or operational security workflows.

The reported distinction is technically material. Offensive tasks can often center on code understanding, vulnerability discovery, and exploit development. Defensive operations also require long-horizon correlation across endpoint, identity, cloud, network, and application data, together with evidence preservation and controls around automated response. In comparable security-automation deployments, evaluation typically depends on false-positive rates, time to containment, analyst override behavior, and the safety of actions taken against production systems, not detection rates alone.

SiliconANGLE reports that Corma began deployments six weeks before the announcement and that its agents are running at Fortune 100 and Fortune 500 organizations in healthcare, financial services, energy, critical infrastructure, and retail. The company reports that early customers reduced threat-response times by more than 94% and expanded security coverage by 15x, according to SiliconANGLE and The Next Web. Those performance figures are self-reported in the cited coverage.

VentureBurn reports that Corma intends to use the funding to expand its interdisciplinary research team and invest in pre-training, post-training, and cybersecurity research. The Next Web describes the team as including frontier AI researchers from Google and DeepMind alongside cybersecurity specialists from Israel's Unit 8200.

For ML and security practitioners, the startup's premise reflects an increasingly important technical question: whether general-purpose models can reliably operate over enterprise security telemetry, or whether security-specific training, tools, and evaluation environments produce meaningfully better defensive behavior. Public evidence of reproducible benchmarks, deployment guardrails, and independently validated operational outcomes would be important measures for assessing systems built around that premise.