Please ensure Javascript is enabled for purposes of website accessibility

Corma Raises $125M to Build Cybersecurity Foundation Models

picture of shield for cybersecurity foundation models

Corma.ai, a startup with offices in San Francisco and Tel Aviv, has raised $125 million to train cybersecurity foundation models aimed at one specific problem: AI is getting better at attacking systems than at defending them. As first reported by SiliconANGLE, the Series A round was led by NEA and values the company at $1 billion.

Returning investors Khosla Ventures, Coatue and Sequoia Capital also took part. The money will go toward more infrastructure and further development of Corma’s models.

Key Takeaways

  • Corma.ai raised $125 million to develop cybersecurity foundation models that strengthen defenses against AI-driven attacks.
  • Current general-purpose models struggle to detect their own cyberattacks; Corma aims to address this issue through specialized models.
  • Corma’s models enhance existing security tools by analyzing telemetry and providing remediation guidance, speeding up breach response significantly.
  • The company’s approach combines automated investigation with human oversight, using a Slack chatbot for easy communication.
  • Corma faces high demand but must expand computing capacity quickly to sustain growth and meet client needs.

The attacker’s advantage Corma wants to close

lock displaying cybersecurity foundation models

Corma’s pitch rests on an uncomfortable imbalance. The company says general-purpose large language models tend to be stronger at launching cyberattacks than at cleaning up after them. To make the point, it ran an experiment in which several LLMs were told to break into a simulated corporate network.

The models managed to plant a malicious file in 85% of the runs. When asked to find the malicious files they had planted themselves, they caught just 12%.

That is Corma’s own test, so it deserves the usual caution. Still, the logic is easy to follow. If the same class of model can break in reliably but can’t spot its own handiwork, defenders who lean on those tools may be handing attackers a head start. Cybersecurity foundation models are the company’s answer to that gap.

What cybersecurity foundation models do in practice

Corma’s software doesn’t replace the security stack a company already runs. It reads the output of those tools. The platform can ingest telemetry from firewalls, network monitoring software and a range of other applications, then look for signs of a breach across that data.

Finding a threat is only half of it. The company says its cybersecurity foundation models also write remediation guidance for each threat they identify, so an analyst isn’t left with an alert and a blank page.

Corma claims some breach response workflows run up to 100 times faster with its models, and that costs drop as well. The company hasn’t been specific in the material available about which workflows those are, so buyers will want to ask.

Why automated breach response is hard

Anyone who has worked in a security operations center knows the problem isn’t a shortage of alerts. It’s too many of them. Automated systems have to sift through a flood of false positives and duplicate notifications before they can reach anything worth investigating.

Data gaps make it worse. Models often see only part of the picture around a breach, partly because many organizations delete security logs after a short period to hold down storage bills. An investigation that starts weeks after an intrusion may be missing the very records it needs.

These are the conditions cybersecurity foundation models have to work in. A model that reasons well on clean benchmark data but stumbles on noisy, incomplete logs won’t help much in a real incident.

Autonomy, with humans in the loop

Corma says its cybersecurity foundation models can run a breach investigation from start to finish without a person involved. When a case needs human judgment, the system alerts the relevant member of the security team.

Staff can also steer the models through a Slack chatbot. The source gives one example: an administrator could ask Corma to double-check whether an earlier attempt to remove malware from a compromised server actually worked. That kind of follow-up question is routine work for security teams, and it’s the sort of task that eats up hours.

The design is sensible. Fully autonomous investigation sounds appealing until the system gets something wrong, and a clear path to a human is how most security leaders will want to adopt it.

Benchmarks and bragging rights

Corma compared one of its cybersecurity foundation models with competitors using DefenseBench V5, an evaluation that scores how reliably a model completes defensive security tasks and how quickly it finishes them. According to the company, its model beat GPT-6 Astra, Claude Opus 5.5 and several other leading LLMs.

Vendor-run benchmark results are a starting point, not a verdict. Security teams considering a purchase should look for independent testing, or run a proof of concept against their own telemetry and incident history.

Demand is outrunning capacity

Chief Executive Alon Pluda, a Corma co-founder, said demand from very large organizations has been overwhelming since the company left stealth. He described the toughest challenge as getting enough computing capacity quickly enough to keep pace.

That helps explain the use of the new funds. Buying infrastructure is a large part of the plan, alongside model improvements. Training and serving cybersecurity foundation models is expensive, and a $1 billion valuation after a Series A signals that investors expect the demand to hold up.

A few related pieces worth your time:

What to watch next

Several questions will decide whether this round for cybersecurity foundation models turns into lasting traction. Here are the ones worth tracking:

  • Whether independent evaluators reproduce Corma’s DefenseBench V5 results.
  • How the models perform when logs are incomplete or have already been deleted.
  • How customers describe the balance between autonomous investigation and human oversight.
  • Whether Corma can add capacity as quickly as new customers arrive.

For business and technology leaders, the broader takeaway is that general-purpose chatbots aren’t the only way to bring AI into security operations. Purpose-built cybersecurity foundation models are now drawing serious money, and Corma’s round is a strong sign that investors think defenders need tools tuned to their side of the fight. Whether cybersecurity foundation models can deliver on the speed and cost claims outside a controlled test is the thing to find out next.

Source: SiliconANGLE

Subscribe

* indicates required