security
Corma CEO shares breakthrough approach to closing the defense gap with AI-powered agents
CEO Alon Pluda of cybersecurity startup Corma explained that the company is creating a “one‑ring” solution to empower defenders with the same power that offensive security researchers wield.
Pluda recounted an incident in which a customer’s security executive was walking his dog when an alert appeared on his smartwatch, generated by a Corma agent.ética The message read, “A live attack has been detected. Please confirm before blocking.” The executive approved the action and the agent halted the malware before it could spread across the company’s network, mitigating the breach in less than ten minutes.
The incident has been described by the customer as “one of the most magical moments of the year,” according to Pluda.
Founded just over a year ago, Corma derives its name from the Elvish word for “ring.” “Our goal is to create a single, all‑powerful ring—this time for defenders,” Pluda said.
Earlier this week, Corma announced a $60 million seed round led by Sequoia Capital, with participation from Khosla Ventures and Coatue. The firm is collaborating with Fortune 100 companies, training models designed to achieve what Pluda calls “superintelligence for defensive cybersecurity.”
Current generative models from OpenAI, Anthropic, and Google excel at coding and language tasks, including bug detection and multi‑step workflow orchestration, but they tend to overlook non‑code defensive work. “Most defensive tasks involve logs, events Designed,” Pluda noted. “That’s not what these models have learned to read reliably.”
To evaluate the capabilities of leading models, Corma tested Claude Opus 4.8, GPT‑5.5, Grok 4.3, and DeepSeek V4 both as attackers and defenders on a simulated enterprise environment. Attackers planted a backdoor, while defenders sought to identify and neutralize it.
Closing the Defensive Gap_STANDARDNESS
Across 241 engagement tests, the models successfully implanted a backdoor in 85 percent of cases but detected only 19 percent of those attacks. “This illustratesन्तकीthe imbalance we’re addressing,” Pluda said. “Offensive capabilities are rapidly improving, but defensive performance lags.”
Corma attributes this gap to the data and objectives that shape general foundation models, which prioritize offensive scenarios. Defensive security, by contrast, requires processing structured machine data—logs, configurations, audit trails, and on‑disk state—and often involves open‑ended reasoning, unlike the clear win-lose objectives of offensive work.
Agentic Defenders
“Defensive security is about finding needles in a haystack,” explained Pluda. Corma’s agents operate like “team members,” capable of handling a range of security tasks across an organization.
Deployments in healthcare, financial services, energy, critical infrastructure, and retail sectors report a 94 percent reduction in threat response times, a fifteen‑fold increase in security coverage, and the discovery of multi‑stage attack campaigns.
“When AI can consistently make the right decisions—understand the domain, optimize the right metrics, and be trusted end‑to‑end—you shrink the window of exploitation dramatically,” Pluda concluded. “It enables teams to cover far more ground than human-only defenses.”
