Corma (AI lab)

Corma is an Israeli-American artificial intelligence startup, headquartered in Tel Aviv and San Francisco, that trains foundation models for defensive cybersecurity and deploys them as AI agents that work alongside corporate security teams.^[1][2][11] Founded in 2025 by chief executive Alon Pluda, the company came out of stealth on August 10, 2026 with $60 million in seed funding led by Sequoia Capital, with participation from Khosla Ventures and Coatue.^[1][3] Corma describes itself as "the first frontier AI lab for defensive cybersecurity" and sells its agents as an AI security "workforce", which it says is deployed at Fortune 100 and Fortune 500 organizations in healthcare, financial services, energy, critical infrastructure and retail.^[3] Its valuation has not been disclosed.^[1]

The company's founding argument, which it calls the "defensive gap", is that general-purpose frontier models have become strong at offensive security because finding and exploiting vulnerabilities is largely a code-reasoning problem, while defense depends on reading logs, telemetry and configuration data that those models handle poorly.^[5][6] Corma is not related to Corma Inc., a Canadian maker of corrugated plastic pipe production systems, or to Corma, a Paris-based software-management startup that raised a €3.5 million seed round in December 2025.^[22][23]

History

Founding and stealth period

Corma was founded in 2025; ISMG gave the founding month as September 2025 and reported that Pluda has led the company since its inception.^[4] SiliconANGLE referred to the company as Corma Labs Ltd.^[9] The name comes from J. R. R. Tolkien's The Lord of the Rings: according to the company, "corma" is the word for ring in Quenya, an Elvish language of Middle-earth, as in the One Ring.^[5] Pluda told Fortune that Corma's product is like that ring, "but this time for the defenders."^[1]

The company raised its seed round while still in stealth; Pluda told Calcalist that the round closed in early 2026.^[2] According to Fortune, Corma deployed its first model to customers about six weeks before its public launch.^[1]

Launch

The round was announced on August 10, 2026 in a Corma press release, a Fortune exclusive and a post by Corma's Sequoia partners, and on the same day Corma published an introductory essay and a research report on the defensive gap.^{[1][3][5][6][7]} The launch was covered by outlets including Calcalist, SiliconANGLE, SecurityWeek, Ynetnews and ISMG.^{[2][9][10][11][4]} Calcalist reported at the time that the company employed 20 people in Tel Aviv.^[2] SiliconANGLE described the seed as Corma's first disclosed outside funding.^[9]

People

Alon Pluda

Pluda is Corma's co-founder and chief executive.^[3] ISMG reported that he spent more than six years in Israeli military intelligence, ending in a role as a cybersecurity specialist.^[4] Mako, which described him as 31 years old in August 2026, reported that he graduated from the Psagot program and served on a vulnerability research team in Unit 8200 before his discharge in 2022.^[12]

After leaving the military he was one of the founders of the cybersecurity company Miggo Security (מיגו סקיוריטי in the Hebrew-language reports), where Calcalist says he was chief technology officer.^[13] Mako reported that he left soon after that company was founded, following a falling-out with the other founders; Calcalist reported that he sued the company and that the two sides reached a settlement shortly after his departure.^[12][13] Mako also reported that Pluda had advised AI startups, including Decart.^[12]

Team

Corma says its team combines AI researchers who previously worked at Google and DeepMind with cybersecurity specialists from Israel's Unit 8200 and from large cybersecurity companies, with experience across pre-training, post-training, and offensive and defensive security.^[3] Hebrew-language coverage said Pluda founded the company together with a founding team of AI and cybersecurity experts, without naming them.^[12] Corma's press release and Sequoia's announcement name only Pluda as a founder.^[3][7]

Hofit Bata lists Corma as her current employer on LinkedIn, and the startup database Raising.fi names her as a co-founder alongside Pluda.^[20][21] Bata was previously a researcher at AI21 Labs^[20] and is the third-listed author of the 2024 paper introducing Jamba, AI21's hybrid Transformer-Mamba language model.^[18] Earlier she co-authored "The Depth-to-Width Interplay in Self-Attention", a NeurIPS 2020 paper with Amnon Shashua and others on how to divide a self-attention network's capacity between depth and width.^[19]

Daniel Ohayon, a Corma researcher who is also affiliated with the Technion, is the lead author of a NeurIPS 2026 paper on efficient attention (see Research).^[17]

Technology

The defensive-gap thesis

Corma's launch materials argue that the areas where frontier models have advanced fastest (writing and refining code, finding and fixing bugs, and orchestrating tools across multi-step workflows) map directly onto offensive security, and that when combined with agentic execution they let models carry out end-to-end attack chains.^[3][5] Both the press release and the company's introductory essay cite Anthropic's disclosure about Claude Mythos as a demonstration.^[3][5] Defensive work, in Corma's account, consists of analyzing large volumes of audit logs, events, configurations and network flows, connecting weak signals over long periods, and making thousands of consistent decisions.^[5]

Pluda told Fortune that most defensive work "doesn't have anything to do with coding" and is instead about "looking at logs, audits, [and] finding the needle in a haystack."^[1] He told ISMG that the company's main goal is repeatability: "If you run it 100 times, you want to get 100 times the same result."^[4] Sequoia's announcement put the same point in training terms, arguing that defense is "out of distribution" for general-purpose frontier models because neither the data nor the intuition it requires is well represented in their training.^[7]

Model training

According to Pluda, Corma's pre-training exposes its models to data modalities and knowledge that are underrepresented in the training data of general-purpose models. Post-training includes reinforcement learning across different security tools and tasks, and the company also uses adversarial training in realistic enterprise environments, where models act against each other as attacker and defender.^[4] Sequoia described cybersecurity as a two-player, zero-sum game with a clean reward, comparable to Go or chess, and said Corma runs large-scale reinforcement learning across environments that replicate real enterprise networks "with all their tools, telemetry and noise."^[7] Sequoia also argued that owning its model weights frees Corma from restrictions that closed labs place on cybersecurity-related use, and that a specialized model costs less per token to run.^[7]

Corma's website says the model can be deployed fully on-premises and post-trained on a customer's own security data.^[15] The company's introductory essay lists the research problems it intends to work on, including rethinking model architecture, teaching models to read "the data and modalities of defense", extending context length "by orders of magnitude", adversarial self-play, and continuous learning so that models can establish baselines over months and years.^[5] As of October 2026, the company's website and blog did not give the model a public name or disclose its size or architecture.^[15][5]

The Corma workforce

Rather than selling a conventional software product, Corma deploys its model as AI agents that carry out tasks end to end across an organization's existing security tools and processes.^[2] Pluda told ISMG that the only thing the company offers enterprises is an "end-to-end security workforce". The agents operate existing security and IT tools and work with human analysts through Microsoft Teams, Slack and ticketing systems, where they can document findings, notify staff and escalate issues.^[4] Hebrew-language coverage reported that the company calls these agents "General Security Employees".^[12] Pluda told Calcalist: "We don't replace anyone and we are not a product; we provide services," adding, "We sell virtual human resources," with each customer deciding how many it needs.^[2]

Pluda told ISMG that customers typically start by delegating repetitive work and give the system more autonomy as they gain confidence. Depending on the customer, the agents either ask for human approval before acting or respond directly, for example by quarantining hosts. He listed network security, cloud security, security operations, threat hunting, incident response, identity and access management, and data security as areas where the company evaluates its system.^[4]

Corma says that in early deployments its agents cut threat response times by more than 94 percent, expanded security coverage by 15 times across security functions, and found multi-stage attack campaigns that would otherwise have gone undetected.^[3] Sequoia's announcement described one customer case in which Corma's agent uncovered, contained and remediated an active attacker campaign within its first hour on the job, a campaign the customer's security team had missed for 52 days.^[7] Ynetnews noted that these performance figures were provided by Corma and had not been independently confirmed.^[11]

Research

Defensive-gap study

On August 10, 2026, Corma published a study that ran frontier models as both attackers and defenders on the same simulated corporate network.^[6] The test range, a fictional holding company called "Kestrel Holdings Group", had 8 subsidiaries, each with its own Active Directory domain, spread across 64 subnets with roughly 349 servers and 4,483 workstations. The range was monitored by Splunk and OpenSearch security information and event management (SIEM) systems, with osquery on Linux and Sysmon and PowerShell logging on Windows.^[6]

In each engagement an attacker model was told to plant a persistent, hidden backdoor; a defender model, told only that a backdoor had been planted at some point, then had to find it. Both sides worked through a single shell tool with the same turn budget. Corma scored results against a before-and-after snapshot of each machine's disk state rather than the agents' own reports, and counted an attack as successful only if the implant survived a reboot.^[6] The study tested Claude Opus 4.8, GPT-5.5, Grok 4.3 and DeepSeek V4 in every attacker and defender pairing, for 241 scored engagements.^[6]

Source: Corma, "Measuring the AI offense-defense gap on a live network".^[6]

The report concluded that whether a defender caught an implant depended mainly on how ordinary the artifact looked (its name, path and surroundings) rather than on the sophistication of the technique. In 26 percent of missed runs, the defender's own commands had surfaced the planted artifact, which the defender then classified as routine automation.^[6] Corma offered an explanation it labeled an interpretation rather than a measured result: defense involves machine data that is a small share of model training data, and calls for open-ended, abductive reasoning, while offense is bounded and goal-directed.^[6]

The headline numbers in Corma's press release differ from those in the report. The press release, and news coverage including Calcalist's, said that across hundreds of simulations AI attackers succeeded 88 percent of the time while AI defenders detected 12 percent of attacks.^[3][2] The report published the same day gives 85 percent and 19 percent.^[6]

Block Sparse Flash Attention

Corma researcher Daniel Ohayon is the lead author of "Block Sparse Flash Attention" (BSFA), written with co-authors affiliated with the Technion. The paper was first posted to arXiv in December 2025; its revised version of September 29, 2026 lists Ohayon's affiliations as Corma and the Technion and states that the paper was accepted to NeurIPS 2026.^[17] BSFA is a training-free sparse attention method for speeding up long-context inference. It computes exact query-key scores inside FlashAttention's tiled framework, then skips the value-side computation for blocks whose maximum scores fall below calibrated thresholds.^[17] On Llama-3.1-8B the authors report up to 1.13 times faster end-to-end inference on LongBench with a 1.1 percent accuracy drop, and up to 1.24 times faster needle-in-a-haystack retrieval at 99 percent of baseline accuracy; the attention kernel alone ran up to 1.38 times faster.^[17] In announcing the acceptance, Corma said reducing latency over long contexts without losing needle-in-a-haystack ability matters for defensive cybersecurity.^[16]

Commentary on AI security incidents

Corma's blog and LinkedIn posts have used AI security incidents to argue its thesis. A September 14, 2026 post reconstructed Anthropic's account of an incident in which an early Claude Opus 4.6 checkpoint, running a misconfigured capture-the-flag evaluation, compromised a real third-party system.^[14] A company post about the OpenAI-Hugging Face agent incident argued that an attacker has a clear success condition while a defender can never know it has not missed something.^[16]

Funding

Sources: Fortune, Calcalist and Corma's press release.^[1][2][3]

Fortune reported that the announcement did not include Corma's valuation.^[1] Sequoia's portfolio page lists Shaun Maguire, Sonya Huang and James Flynn as its partners on the company, and the firm's announcement was written by those three together with Carl Eschenbach.^[8][7] A LinkedIn post announcing the round, reposted by Corma's company page, also thanked Lia Weiner-Tsur of Netz Capital alongside partners at Sequoia, Khosla Ventures and Coatue.^[16] According to Fortune, Corma planned to use the money to scale its models by expanding the data and training behind them, and to hire in defensive security, AI and research.^[1]

In the launch announcement, Maguire said Corma "has trained its model for the complexity of real-world attacks and is building the intelligence layer defense actually needs." Vinod Khosla said that as cyberattacks become more autonomous, "the stakes move beyond data and finances to critical infrastructure, healthcare systems, and essential services."^[3]

Reception

Ynetnews pointed out that both the simulation results and the customer performance numbers came from Corma and had not been independently verified.^[11] Debarghya Das's "The Ultimate Neolab List", which catalogs new AI labs, includes Corma but rates it a "Borderline" case, noting that it trains its own models with reinforcement learning but is "a vertical security vendor already selling to enterprises."^[24]

See also

• List of neolabs
• Neolab
• AI in Cybersecurity
• CrowdStrike SafeMind
• Project Glasswing
• Claude Mythos Preview

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Emily Forlini, Fortune. "Exclusive: Corma raises $60 million from Sequoia for AI trained to defend against cyberattacks." August 10, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Meir Orbach, Calcalist (Ctech). "Corma raises $60 million Seed to build AI designed to fight AI-powered cyberattacks." August 10, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Corma, via ACCESS Newswire. "Corma, the First Frontier Defensive Cybersecurity AI Lab, Raises $60M as AI Supercharges Attackers." August 10, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Michael Novinson, GovInfoSecurity (ISMG). "Corma Raises $60M to Build Cyber Defense Foundation Models." August 12, 2026.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Corma. "Introducing Corma." August 10, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Corma. "Measuring the AI offense-defense gap on a live network." August 10, 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Shaun Maguire, Sonya Huang, James Flynn and Carl Eschenbach, Sequoia Capital. "Partnering with Corma: Closing the Defensive Cybersecurity Gap." August 10, 2026.
8. ^Sequoia Capital. "Corma." Portfolio page, accessed October 27, 2026.
9. ^1 ^2 ^3Duncan Riley, SiliconANGLE. "Corma launches with $60M in funding for defensive cybersecurity AI." August 10, 2026.
10. ^SecurityWeek News. "Corma Raises $60 Million for Defensive Cybersecurity AI Model." August 11, 2026.
11. ^1 ^2 ^3 ^4ynet Global, Ynetnews. "Corma raises $60 million as AI gives cyberattackers a growing edge." August 10, 2026.
12. ^1 ^2 ^3 ^4 ^5Mako (N12). "חברה ישראלית נחשפת עם גיוס של 60 מיליון דולר: מפתחת מודל סייבר הגנתי ראשון מסוגו" (in Hebrew). August 10, 2026.
13. ^1 ^2Meir Orbach, Calcalist. "הסטארטאפ שמפתח 'עובד וירטואלי': גיוס ענק של 60 מיליון דולר לקורמה" (in Hebrew). August 10, 2026.
14. ^1 ^2Corma. "Claude Tried to Quit 8 Times. Then It Hacked a Real System." September 14, 2026.
15. ^1 ^2Corma. "Superintelligence for defensive cybersecurity." Company website, accessed October 27, 2026.
16. ^1 ^2 ^3Corma. "Corma." LinkedIn company page and posts, accessed October 27, 2026.
17. ^1 ^2 ^3 ^4 ^5Daniel Ohayon, Itay Lamprecht, Itay Hubara, Israel Cohen, Daniel Soudry and Noam Elata. "Block Sparse Flash Attention." arXiv:2512.07011, v1 December 7, 2025; v2 September 29, 2026.
18. ^Opher Lieber, Barak Lenz, Hofit Bata, et al. "Jamba: A Hybrid Transformer-Mamba Language Model." arXiv:2403.19887, March 28, 2024.
19. ^Yoav Levine, Noam Wies, Or Sharir, Hofit Bata and Amnon Shashua. "The Depth-to-Width Interplay in Self-Attention." arXiv:2006.12467 (NeurIPS 2020), June 22, 2020.
20. ^1 ^2Hofit Bata. "Hofit Bata - Corma." LinkedIn profile, accessed October 27, 2026.
21. ^Raising.fi. "Corma Secures $60 Million in Seed Funding Led by Sequoia Capital." August 10, 2026.
22. ^Wikipedia. "Corma Inc." Accessed October 27, 2026.
23. ^Corma (Paris), via PR Newswire. "Corma Announces €3.5M Seed Round to Revolutionize Software License and Access Governance." December 8, 2025.
24. ^Debarghya Das. "The Ultimate Neolab List." Accessed October 27, 2026.