Corma is a 2025-founded company, based in Tel Aviv and San Francisco, that trains its own foundation models for one job only: defending computer networks. Its pitch is that the models everyone else builds are better at breaking in than at keeping intruders out — and that defense deserves models trained for defense. Investors, fresh off a month of high-profile breaches, appear to agree.

The round in numbers#

FigureDetail
Series A$125 million
Valuation$1 billion (whether pre- or post-money was not disclosed)
Total raised$185 million
Seed round$60 million, announced August 10, 2026
Lead investorNEA
Returning investorsSequoia Capital, Khosla Ventures, Coatue
Founded2025; offices in Tel Aviv and San Francisco
HeadcountAbout 20 people, according to one report

The speed of the markup is what stands out. A $1 billion valuation three months after a product’s public debut is rare even in a frothy market — and it arrives before the company has published a single revenue figure, customer count, or loss. What investors bought, for now, is a thesis and a team, not a track record.

Why now: the offense–defense gap#

Corma’s argument starts with its own research. In an August study, the company pitted general-purpose AI models against each other inside a controlled enterprise network: the attackers established persistent backdoors in 85% of runs, while the defenders detected only 19% of implants across 241 engagements. The numbers come from Corma’s own experiment in a synthetic environment, so they are not real-world attack rates — but the direction of the gap is the company’s whole case for existing.

October supplied the urgency. A critical flaw in PHP surfaced on October 21, the same day flaws and incidents were disclosed in Node.js and LiteLLM, the popular proxy for routing requests to language models. Chief information security officers spent the month scrambling, and, according to TechCrunch, interest in Corma’s product jumped. NEA has taken to calling the stretch “Black October.” It is the backdrop the round was sold against: the week AI security stopped being a conference topic and became a board-level problem.

A luminous guardian figure holding a shield between a corporate network and a swarm of incoming red threat signals
AI-generated illustration. Corma’s pitch: a dedicated defender standing between enterprise networks and automated attacks.

What Corma actually sells#

The product is a set of foundation models trained from scratch for defensive cybersecurity, deployed as autonomous agents — an “AI workforce” — that works inside a customer’s existing tools alongside the human security team. The models take in telemetry from firewalls, network monitoring software and other security tools, look for breaches, write remediation guidance, and run investigations on their own, alerting staff when a human decision is needed.

Building the models from scratch, rather than layering a product over systems from OpenAI, Anthropic or DeepSeek, is the point of differentiation Corma leans on hardest. The company says its server capacity has grown 24-fold since August, and the new money goes to more infrastructure and better models. Its customers reportedly include Fortune 100 and Fortune 500 companies in healthcare, financial services, energy, retail and telecommunications, with annual recurring revenue estimated in the tens of millions of dollars — figures reported by the Israeli press, not confirmed by the company.

The caveats: a valuation is not a verdict#

Every number that would let an outsider judge the business is missing. Corma has published no revenue, no customer count, and no losses. The performance figures on its product page — 99% successful autonomous task completion, 100 times faster end-to-end completion, 95% lower cost per completed task — are vendor claims with no published methodology behind them, and should not be treated as independently verified benchmarks.

The same caution applies to DefenseBench V5, the evaluation in which Corma says its model outperformed GPT-6 Astra and Claude Opus 5.5. It is a benchmark Corma built itself, and no independent results have been published. And the field is crowded: Kevin Mandia’s new startup Armadin is pursuing the same defensive-AI idea, and Anthropic, OpenAI and Microsoft are all building defensive capabilities of their own. A $1 billion price tag three months after launch says more about investor appetite for AI security than about proven results.

A glowing circuit-board chart climbing steeply toward a bright horizon above rows of server racks and shields
AI-generated illustration. The valuation curve, so far, is steeper than the evidence curve.

What to watch#

  • Disclosed revenue and customers. The first hard numbers will say more about Corma than any valuation. Watch for named customers and a revenue figure the company itself stands behind.
  • Independent evaluation. The offense–defense gap and the DefenseBench results both come from Corma. Third-party testing of its models against general-purpose frontier models on defense tasks is the missing check.
  • Whether the own-models thesis holds. Corma’s bet is that purpose-built defense models beat repurposed general-purpose ones. As frontier models get better at defense, that gap may narrow — or widen.
  • Conversion of the October surge. The breach-driven interest that fed this round needs to turn into signed, renewing contracts rather than pilots.

For now, Corma is the clearest price tag yet on a shift the whole industry felt this autumn: AI is no longer just the thing security teams defend against. It is the thing they are buying — and investors are paying up to own the company that builds it for defense first.

Sources: TechCrunch, “Cybersecurity AI startup Corma reaches $1B valuation in about 3 months”; Business Wire, “Corma Raises $125M Series A at $1B Valuation to Scale Its Frontier Foundation Model for Cyber Defense”; SiliconANGLE, “Corma raises $125M to train cybersecurity-optimized foundation models”.