Corma Labs

**Product and problem.** Corma is building a specialized defensive-cybersecurity AI workforce rather than another point product for endpoint, identity, or vulnerability management. The company says its foundation models work alongside human security teams, generalize across defensive tasks, and operate through the security stack that a customer already uses. The concrete problem is an asymmetry in cyber operations: automated offense can search, exploit, and persist at machine speed, while defense still depends heavily on people correlating logs, events, configurations, alerts, and business context under time pressure. Corma’s proposed answer is a model that can perform more of the investigative and operational workload end to end, helping a small security team cover more systems and incidents without requiring every workflow to be redesigned around a new console. Public materials do not enumerate the complete product surface, supported tools, or customer-specific action permissions, so the record should treat “AI workforce” as the company’s product framing rather than as proof of unrestricted autonomous remediation.

**Core technology and operating model.** Corma publicly describes a foundation model built from the ground up for defensive cybersecurity. Its materials emphasize that defensive work is not simply code generation: it involves interpreting noisy telemetry, forming and testing hypotheses, finding anomalous behavior in a large normal environment, and deciding whether and how to act. The model is intended to power agents that can use existing security tools and assist with investigation, containment, and remediation. Sequoia’s announcement describes Corma’s training direction as applying reinforcement learning and self-play to a two-player cyber-defense setting, but that account comes from the lead investor and the company has not publicly disclosed model size, training corpus, data-rights posture, reward design, or independent benchmark methodology. Corma claims materially stronger defensive performance and much lower cost and latency than general-purpose models. Its stated deployment options include fully on-premises operation and post-training on a customer’s security data and context. Those are strategically important capabilities, but diligence still needs reproducible task-level evaluations, false-positive rates, tool-use guardrails, audit trails, and evidence that local adaptation improves outcomes without leaking sensitive information.

**Market, customers, and go-to-market.** The initial market is enterprise security operations, where CISOs already purchase telemetry, detection, response, identity, cloud, and managed-security products but remain constrained by analyst capacity. Corma’s “works where the human team already works” positioning suggests an integration-led enterprise motion: add specialized agents to existing workflows and data sources, then expand as the workforce proves useful across more defensive tasks. CTech reported that the company was already working with Fortune 100 and Fortune 500 organizations, while Sequoia described engagements across healthcare, finance, critical infrastructure, retail, and other large enterprises. The official site says important organizations already use the workforce, but it does not name customers, publish contracts, or disclose revenue, retention, deployment counts, or average time to value. Public evidence therefore supports early enterprise access and credible buyer interest, not yet repeatable product-market fit. Procurement will depend on least-privilege integrations, security review, data residency, human approval policies, and a clear economic case against adding more analysts or buying bundled AI from an incumbent platform.

**Traction, financing, and third-party validation.** Corma was founded in 2025 and emerged publicly in August 2026 with a $60 million seed round led by Sequoia Capital, with Khosla Ventures and Coatue participating, according to CTech and Fortune. Startup Nation Finder lists the same financing, an August 2026 date, approximately 20 employees, and Alon Pluda as founder; CTech likewise reported 20 employees in Tel Aviv. Sequoia’s portfolio record marks the company as founded in 2025 and partnered in 2026, and its partner announcement confirms that the fund is backing Corma’s defensive-cybersecurity foundation-model thesis. Corma has also published a research report comparing frontier-model offense and defense in a controlled enterprise range. That report describes 241 scored engagements and reports persistent backdoors in 85% of attack runs versus detection of 19% by general-purpose defenders; it is useful evidence of a serious evaluation program, but it is company research and should not be treated as an independent product benchmark. Investor-reported customer anecdotes are encouraging, yet named references and independently audited operating metrics remain absent from the public record.

**Founders and team.** Alon Pluda is publicly identified as Corma’s founder and CEO by Sequoia and CTech. The available public record presents the team as an interdisciplinary combination of advanced-AI researchers and highly capable cybersecurity practitioners, which is the right talent mix for a company attempting to train models around operational defense rather than merely fine-tune a chat interface. However, the public company page names only Pluda, and broad biographical claims about other employees, prior employers, military units, or individual research accomplishments are not sufficiently documented in the sources used for this record. The reported Tel Aviv team of about 20 people is a meaningful early hiring base, with the company also maintaining a San Francisco presence, but it is too small to assume mature 24/7 support, evaluation, security engineering, sales engineering, and government-procurement capacity. Team diligence should verify who owns model research, cyber-range design, customer deployment, trust and safety, and enterprise integrations, as well as whether the company can retain scarce security and frontier-model talent after a highly visible seed round.

**Competitive dynamics.** Corma competes with several different categories rather than one direct peer. CrowdStrike, Microsoft Security, Palo Alto Networks, Google Security Operations, and SentinelOne all have large telemetry footprints and can add increasingly capable AI assistants to established detection and response platforms. Tines competes through security workflow automation, while AI-native SOC vendors such as Prophet Security and Dropzone AI compete for analyst augmentation and investigation budgets. Corma’s possible advantage is the combination of a model trained specifically for defensive cyber reasoning, agents that generalize across a customer’s existing tools, and sovereign deployment for organizations that cannot place sensitive security context in a standard multitenant service. The tradeoff is execution risk: incumbents have distribution, data, integrations, and procurement trust, while narrower AI competitors may move faster on a defined SOC workflow. Corma must show that a vertically trained model produces better defended environments and lower total operating cost, not simply stronger marketing language. Its likely competitive edges are (1) domain-specific training objectives, (2) end-to-end workforce behavior, and (3) on-premises adaptability.

**Defense, security, and resilience relevance.** Dual-use relevance is direct because cyber defense is simultaneously a commercial necessity and a national-security mission. The same agents that investigate identity abuse, cloud compromise, malware persistence, or suspicious network activity for a bank or hospital could support government networks, defense contractors, communications providers, energy operators, and other critical infrastructure. On-premises deployment and customer-specific post-training are especially relevant to sovereign, classified-adjacent, or data-residency-constrained environments where telemetry cannot be sent to a public cloud. Corma’s own defensive-gap research also addresses a strategically important question: whether general AI is scaling offensive capability faster than defensive capability. That supports the resilience thesis, but it does not prove fielded military capability. No public source used here establishes a defense contract, classified authorization, government customer, or certification. The responsible strategic interpretation is therefore a high-relevance defensive technology with plausible deployment into defense and critical infrastructure, subject to accreditation, isolation, supply-chain, model-governance, and human-override requirements.

**Stage, trajectory, and diligence risks.** Corma is best classified as early stage: it has a 2025 founding date, a very large seed financing, a reported 20-person Tel Aviv team, and evidence of early enterprise work, but limited public operating history. The financing gives it unusual runway to train models, build cyber-range evaluation infrastructure, and recruit scarce talent. The principal diligence risks are substantial: (1) public performance claims may not transfer from controlled ranges to messy production networks; (2) an agent with access to security tools can cause material damage through hallucination, overreaction, or a compromised integration; (3) training and post-training on security data raises privacy, provenance, retention, and cross-tenant isolation questions; (4) model inference and on-premises deployment may be expensive to operate at enterprise scale; (5) Microsoft, Google, CrowdStrike, Palo Alto Networks, and other incumbents can bundle adjacent capabilities; and (6) public customer, revenue, retention, safety, and certification evidence is still thin. The trajectory is compelling, but the next proof points should be independent efficacy tests, production references, measurable analyst-time savings, renewal data, safe-action controls, and a credible path through regulated and sovereign procurement.