Analysis
The Round
Corma emerged from stealth on August 10 with $60 million in seed funding led by Sequoia Capital, joined by Khosla Ventures and Coatue, to build what it calls the first frontier foundation model purpose-built for defensive cybersecurity, according to Fortune. The company, founded in 2025 and based in Tel Aviv and San Francisco, has already deployed its first model to Fortune 100 and Fortune 500 organizations across healthcare, financial services, energy, critical infrastructure and retail, according to Calcalistech.
The Problem Corma Says It's Solving
Corma's pitch is a specific asymmetry: AI's offensive cybersecurity capabilities have been advancing fast enough that OpenAI just shipped a model answering 95% of advanced exploit-development prompts for vetted defenders, while general-purpose defensive tooling has lagged behind. Rather than adapting an existing consumer or coding model, Corma built a model from the ground up trained specifically on defensive tasks -- threat detection, incident triage, security-operations workflows -- the same specialization bet that has driven Horizon3's and Zenity's recent rounds in adjacent corners of AI security.
The Competitive Field
Corma enters a crowded and fast-moving field. Horizon3 raised $250 million at a $2 billion valuation earlier this month for autonomous security-validation testing, and Zenity raised $125 million to secure AI-agent deployments specifically. What differentiates Corma's pitch is scope: rather than a point solution for one part of the security stack, the company is positioning itself as a foundation-model layer that other security products could build on top of, similar to how OpenAI and Anthropic sell general-purpose models that get fine-tuned for narrower tasks downstream.
Numbers in Context
A $60 million seed is a large check even by 2026's inflated AI-funding standards, and Sequoia, Khosla and Coatue all writing seed checks to the same six-week-old deployment signals investor conviction that defensive AI is under-capitalized relative to how fast offensive AI capability is moving -- a gap OpenAI's own Astra disclosure this month made concrete.
Why This Round Landed the Same Week as OpenAI's
Corma's seed closed the same week OpenAI shipped GPT-5.6-Cyber, a purpose-trained model for vetted defenders that answered 95% of advanced exploit-development prompts in internal testing versus 1.5% for its consumer counterpart. That's not a coincidence of timing so much as two different bets on the same underlying trend -- a frontier lab betting a gated version of its general-purpose model is the right defensive tool, and a venture-backed startup betting a model trained from scratch on defense specifically will out-perform a fine-tuned general-purpose one. Which approach wins is an open, testable question over the next few product cycles. Corma's pitch to enterprise buyers is specifically that a model never trained to be a helpful, broadly capable assistant in the first place won't carry the same refusal instincts a fine-tuned general model has to be coaxed past, a training-from-scratch advantage that's easier to claim in a pitch deck than to prove in a customer bake-off.
The Counterweight
Six weeks of Fortune 500 deployment is an early signal, not proof of product-market fit, and "first foundation model purpose-built for defensive cybersecurity" is a marketing claim Corma hasn't had to defend against a head-to-head benchmark yet. The defensive-AI category is also getting crowded fast enough that differentiation on trained-from-scratch specialization alone may not hold if a general-purpose lab like OpenAI or Anthropic decides to fine-tune its own frontier model for the same use case at a fraction of Corma's training cost.
Ahead
The real test is whether Corma's model actually detects and stops attacks its Fortune 500 customers' existing tooling missed -- a claim that will be provable or disprovable within the next few incident-response cycles, not on a pitch deck.
