Key Takeaways
- Corma’s AI agents act as autonomous “team members” that can detect and block live attacks with minimal human intervention.
- Frontier language models are far better at offensive security tasks (exploit creation, vulnerability discovery) than at defensive security work.
- The “defensive gap” stems from models being trained mostly on prose and code, while defense relies on structured machine data and open‑ended reasoning.
- In controlled tests, models implanted a persistent backdoor 85 % of the time but detected only 19 % of attacks.
- Early deployments with Fortune 100/500 firms have cut threat‑response times by >94 % and expanded security coverage 15‑fold.
- Corma recently raised $60 million in seed funding led by Sequoia Capital, Khosla Ventures and Coatue.
Introduction
Corma, an AI‑focused cybersecurity startup founded roughly a year ago, aims to give defenders the same level of intelligent power that attackers now enjoy from cutting‑edge language models. CEO Alon Pluda describes the company’s ambition as building “the one ring to rule them all, but this time for the defenders to have this power.” The startup’s core thesis is that while foundation models from OpenAI, Anthropic, Google and others have become superb at offensive security—finding bugs, writing exploit code, and chaining multi‑step attacks—they lag significantly when asked to perform routine defensive tasks such as log analysis, configuration review, and incident response.
Customer Anecdote: The Dog‑Walking Executive
Pluda illustrated the potential of Corma’s agents with a striking real‑world story. A senior security executive at a Fortune 100 client was out walking his dog when his smartwatch buzzed with a notification from a Corma agent: “I just caught a live attack. I need your permission to block it.” After granting approval, the agent isolated the malware, halted lateral movement, and mitigated the intrusion in under ten minutes. The executive later told Pluda that “walking outside with his dog, and blocking a real‑live attack with his AI coworker” felt like “one of the most magical moments of his year.” This episode underscores how autonomous agents can act as trusted teammates, intervening swiftly without waiting for a human analyst to triage alerts.
Founding Vision and the “Defense Gap”
Corma was created to close what Pluda calls the “defense gap”: the growing asymmetry where AI models excel at offense but struggle with defense. He notes that the gap is not merely a matter of model size or compute; it is rooted in the nature of the tasks themselves. Offensive security often has a clear, measurable objective—“make this work” or “break this”—and a concrete finish condition that can be verified quickly. Defensive work, by contrast, involves sifting through logs, events, audit trails, and system states to find subtle anomalies, a process that is far more open‑ended and less amenable to the next‑token prediction objectives that dominate current language‑model training.
Funding and Market Traction
Earlier this week Corma announced a $60 million seed round led by Sequoia Capital, with participation from Khosla Ventures and Coatue. Pluda revealed that the startup is already working with dozens of Fortune 100 and Fortune 500 organizations across sectors such as healthcare, financial services, energy, critical infrastructure, and retail. These early adopters have deployed Corma’s AI workforce—software agents that function like additional security team members—to handle a variety of defensive duties, from continuous monitoring to incident response.
Experimental Evidence of the Imbalance
To quantify the disparity, Corma conducted a head‑to‑head benchmark using four frontier models: Claude Opus 4.8, GPT‑5.5, Grok 4.3, and DeepSeek V4. Each model was tested both as an attacker (tasked with planting a persistent backdoor) and as a defender (charged with discovering and stopping that backdoor) inside a synthetic network mimicking a multi‑business enterprise. Across 241 scored engagements—15 repetitions for every attacker‑defender pairing, including self‑matches—the models succeeded in implanting a backdoor in 85 % of runs. Yet, when acting as defenders, the same models detected only 19 % of the attacks. Pluda emphasized that this stark contrast “speaks to the inherent imbalance we are trying to solve” and confirms that foundation models are improving offensively at an exponential rate while defensive capabilities lag far behind.
Why Defense Is Harder for Current Models
Pluda explained two core reasons for the defensive shortfall. First, the training data for large language models consists predominantly of natural language text and source code; structured machine data such as system logs, registry entries, and network flow records represent a tiny fraction of what the models see during pretraining. Consequently, the models have limited exposure to the patterns and formats that dominate defensive analysis. Second, defensive reasoning is intrinsically open‑ended: analysts must hypothesize about unknown threat behaviors, correlate disparate signals, and decide when an anomaly merits action. Offensive tasks, by contrast, often present a clear success criterion (e.g., gaining privileged access) that aligns neatly with the predictive objectives of current models. Together, these factors create a mismatch between model strengths and the demands of real‑world security defense.
AI Agents as a Generalized Security Workforce
Corma’s solution is to wrap its fine‑tuned models in agentic scaffolding that enables them to perceive, reason, and act across the full spectrum of defensive security tasks. Pluda likens these agents to “team members” that can be assigned to any security function—monitoring SIEM alerts, reviewing configuration drift, hunting for indicators of compromise, or orchestrating response playbooks. Because the agents operate continuously and can scale horizontally, they dramatically increase the amount of ground a security program can cover without proportional growth in headcount. The startup claims that early deployments have reduced threat‑response times by more than 94 %, expanded security coverage by roughly 15 times across different security domains, and uncovered multi‑stage attack campaigns that had previously evaded detection.
Impact and Outlook
The combination of rapid autonomous response and expansive coverage translates into tangible business value: faster containment limits breach costs, broader visibility reduces blind spots, and the ability to trust AI to act decisively frees human analysts to focus on higher‑level strategy and threat hunting. Pluda envisions a future where AI defenders work alongside human analysts in a true partnership, handling the noisy, repetitive work of detection while humans provide contextual judgment and creative problem‑solving. With fresh capital and a growing roster of enterprise customers, Corma is positioned to push the frontier of defensive AI, striving to ensure that the “one ring” of power ultimately serves those who protect rather than those who attack.