Corma, a cybersecurity startup with operations in Tel Aviv and San Francisco, announced a $125 million Series A on November 16, 2026, at a $1 billion valuation. New Enterprise Associates (NEA) led the round, with existing investors Sequoia Capital, Khosla Ventures and Coatue participating. The company says it has raised $185 million in total, including the $60 million seed it disclosed when it emerged from stealth in August. Corma’s funding announcement
The financing will go toward developing larger or more capable models, increasing computing capacity and selectively hiring, according to Corma. CEO and co-founder Alon Pluda started the company in 2025. It has about 25 employees across its operations, according to CTech. For Tel Aviv, the round is a bet on a local team building the models that power its security software as well as the software itself. CTech’s reporting
The job Corma wants its models to do
A security team investigating a suspicious login may have to reconcile an identity alert, a device record and a series of network events before it knows whether to respond. The relevant entries can look innocuous when read alone. An AI system that simply summarizes the alert can save some reading. An investigation tool also has to find supporting evidence, weigh conflicting signals and give an analyst enough information to check the conclusion.
Corma’s agents are designed to use an organization’s existing security tools to investigate incidents and produce guidance on how to respond. SiliconANGLE reports that the software can draw on firewall and network monitoring data, ask security staff for input when needed and take direction through Slack. The way an agent fits into the existing operation matters because a customer’s tools determine which records it can see and what actions it can take. SiliconANGLE’s reporting
NEA describes Corma’s approach as reinforcement learning in enterprise-like environments, where a model can learn from whether it accomplished a task that can be checked. The investor says Corma has deployments at large businesses in sectors including healthcare, financial services and energy. Those accounts come from the investor backing the round. The number of deployments and customer-by-customer results would help readers judge how widely the approach has been tested in production. NEA’s investment note
What the August experiment measured
In an August 10 study, Corma put four general-purpose models in attack and defense roles on an isolated network built to resemble a large company. Across 241 scored engagements, attack agents established a backdoor that survived a reboot in 85% of runs. In the subset with a verified persistent backdoor, defender agents identified it in 22% of cases. The researchers checked the findings against changes on the machines instead of asking each model to grade its own work. Corma’s experiment and scoring method
The lab reset the network between engagements. Both agents had the same tool and turn limit, and the defender’s briefing said an intruder had already gained access and likely left a backdoor. That makes the result a measure of how well the models hunted for a suspected compromise given broad access to a lab network. Study design and prompts
The study helps explain Corma’s choice of problem. It examined one kind of attack after initial access, on a network the company controlled. A security team evaluating its tools would face many other cases, from credential theft to routine alerts that turn out to be harmless. Those cases can have incomplete logs, conflicting records or evidence that arrives hours apart. An evaluation on a customer’s own data would answer a different question about how the product performs in that environment.
What to watch after the round
Corma says it plans to share its next generation of models after putting more computing capacity behind development. That puts attention on how the company will demonstrate progress. A faster investigation is useful when its evidence holds up under review. A convincing evaluation would also measure missed incidents and false alarms, and show how much time staff spend checking the output. Corma’s plans for the new funding
For security teams, a crucial distinction is between letting a tool collect evidence and giving it authority to act. Examining a log and suspending an account have different consequences. The next proof point for Corma will be how much investigative work its models can take on across different customers, and whether those teams can set and enforce the boundaries they need.




