TL;DR
Corma raised a $125 million Series A to build specialized defensive AI agents. Its research and product claims need independent scrutiny.
Introduction
Security teams do not merely need an agent that can find a software bug. They need to distinguish a genuine incident from normal activity across a sprawling collection of identities, devices, logs and cloud services. A wrong action can disrupt a business just as a missed attack can.
Corma’s funding announcement puts a price on the bet that defense deserves its own models. The company says NEA led its $125 million Series A at a $1 billion valuation, with Sequoia Capital, Khosla Ventures and Coatue participating. The round brings total funding to $185 million. (Source : Corma — Three months in)
A valuation reflects investor expectations, not a measured reduction in customer risk. The more useful question is whether specialist agents improve the economics of investigation without introducing unacceptable new failure modes.
A gap worth investigating
Corma emerged from stealth with $60 million in seed funding three months before the Series A. It says it plans to use the new capital to scale its models and compute capacity, and selectively expand its team. The company also says compute supply has been a constraint on serving demand. (Source : Corma — Three months in)
The company’s August research gives context for its thesis. In Corma’s controlled enterprise-network experiment, general-purpose AI attackers established persistent backdoors in 85% of runs, while general-purpose AI defenders detected 19% of implants. The study reports 241 engagements scored against disk differences. (Source : Corma — Measuring the AI offense–defense gap)
Those figures are company-reported results from a specific synthetic environment, not rates for real-world attacks. The attacker began with privileged network access, and both sides operated under the study’s particular tools and turn limits. They do not establish a universal offense-defense gap, nor do they demonstrate that Corma’s product would have detected the missed implants.
Still, the experiment frames a practical distinction. Offensive work in that setup had a concrete target: establish persistence. Defensive work meant identifying a needle in a much larger set of legitimate activity. An agent’s coding ability alone may not settle that second problem.
From model to security workflow
Corma describes its Workforce as using customers’ existing security and IT tools to investigate, connect evidence and take action. Its product page says teams can assign tasks, review findings and approve actions through interfaces including Slack and Teams. (Source : Corma — AI Security Workforce)
That workflow makes integrations part of the product’s risk surface. An agent with permission to inspect alerts can also receive sensitive context; an agent allowed to isolate endpoints can interrupt legitimate work. Buyers should distinguish read-only investigation from write-capable response, and verify how credentials, approval policies and audit logs are handled in their own deployment. The product page’s mention of approvals does not prove that every consequential action requires approval.
The vendor advertises 99% successful autonomous task completion, 100 times faster end-to-end completion and 95% lower cost per completed task. Those are vendor claims; the public product page does not explain the task mix, baselines or measurement methodology behind them. They should not be treated as independently verified benchmarks. (Source : Corma — AI Security Workforce)
What an enterprise should measure
A credible pilot needs a baseline and a denominator. Compare the same kinds of cases with and without the agent, including cases where nothing malicious occurred. Track confirmed detections, missed incidents, false positives, time spent reviewing agent output and time to resolution. Report the severity and mix of cases rather than relying on an aggregate success percentage.
Run response actions behind explicit approval at first, with scoped credentials and a rollback plan. Measure whether a faster investigation merely shifts effort to human reviewers or actually reduces end-to-end workload. Escalation quality matters: a system that confidently closes an ambiguous case may look efficient while increasing risk.
This is the same boundary problem seen in agent payments: tool access is not equivalent to permission to take every action. For a security agent, the protected asset may be an employee’s access or a production server rather than a bank card.
FAQ
What did Corma raise?
The company announced a $125 million Series A at a $1 billion valuation, bringing total funding to $185 million.
Did its study prove that Corma detects more attacks?
No. The cited experiment tested general-purpose models as attackers and defenders in a controlled environment; it does not establish product performance.
Are the advertised productivity figures independently verified?
The product page presents them as vendor claims without the supporting methodology needed to assess them independently.
Should an agent be allowed to isolate systems automatically?
That depends on the deployment and its risk tolerance. A cautious pilot can start with read-only access and require approval before disruptive actions.
Further Reading
- Corma — Series A announcement
- Corma — Offensive and defensive AI study
- Globes — Corma financing and business
- The Agent Report — Human approval for agent payments
— The Agent Report