The same AI that is getting better at writing code is getting better at breaking it. Vulnerabilities are just another kind of bug, and agents can now exploit what they find end to end. Defense is a different problem: it means digging through audit logs, security events and network flows, correlating weak signals across systems and over long time horizons, and staying consistent across thousands of decisions in a row. Corma, a startup that trains its own foundation models for that work, has reached unicorn status three months after emerging from stealth.

Corma trains foundation models for defensive cybersecurity and deploys them as an AI workforce that works alongside enterprise security teams. A foundation model here means a base AI model trained at scale for a specific domain rather than adapted from a general-purpose one. According to Sequoia Capital, Corma trains its model with reinforcement learning on simulated enterprise networks that include their tools, telemetry and noise. The model is not a technology supplied to other security vendors. It powers Corma’s own workforce, which runs inside the security tools a customer already uses and carries out tasks from start to finish. In effect, Corma trains the model and places it with customers as AI staff. CEO Alon Pluda told Israeli outlet Calcalist, “We don’t replace anyone and we are not a product,” and said the company sells virtual human resources.
The company says coding models were never optimized for defensive work, so the progress that makes AI a better attacker barely moves defense. It is headquartered in Tel Aviv and San Francisco, and its team includes researchers from Google DeepMind and security experts from Israel’s Unit 8200.
Alon Pluda is co-founder and CEO. Sequoia Capital, which led Corma’s seed round, has described him as one of the most elite hackers in the world.
Corma announced a $125 million Series A at a $1 billion valuation. NEA led the round, with existing investors Sequoia Capital, Khosla Ventures and Coatue participating. It comes three months after Corma disclosed a $60 million seed led by Sequoia in August, bringing total funding to $185 million.
According to the company, demand has surged across the Fortune 100, Fortune 500 and Fortune Global 500 since it left stealth, and it has expanded server capacity 24x to keep up. It says some large enterprises already had the Corma workforce in place when “Black October,” a wave of critical vulnerabilities and breaches in widely used software, hit in October. Pluda said securing enough compute has been one of the company’s hardest problems. NEA partner Aaron Jacobson said Corma is building the intelligence layer defensive security needs, and called it the only company taking on the problem in a way that scales across organizations. Corma plans to use the new funds to scale its foundation models and compute, and said it will share its next-generation models soon.
Agentic SOC Startups Chase the Same Budget
Corma’s closest rivals are startups automating security operations with AI. 7AI raised a $130 million Series A, which it billed as the largest in cybersecurity history, led by Index Ventures. Torq raised a $140 million Series D and became a unicorn. Exaforce raised a $125 million Series B for its agentic SOC, bringing total funding to $200 million. While they ship agents and platforms that automate security operations, Corma pitches an AI workforce driven by its own model instead of software. The comparison Corma itself draws is with general-purpose frontier models, which it argues are a poor fit for defensive work.
For more on competitors, see here.
MORE FROM THE POST
- Upwind Raises $400M Series C at $3.8B Valuation as Cloud Security Goes Runtime-First
- Mirendil, a Frontier Lab Where Agents Do the AI Research, Raises $1B Series A
- Text-and-Call AI Assistant Instinct’s Valuation Quadruples in a Month, Raises $1B
- Agentic Control Plane Maker Island Raises $400M Series F at $6.4B Valuation
- Factory Raises $200M at $5B Valuation, Tripling in Five Months
Share
Most Read
- 1
- 2
- 3
- 4
- 5


Leave a Reply