Sentra’s $50M Bet: Small AI Models Secure Data Classification

For CISOs grappling with skyrocketing unstructured data and tightening budgets, the prospect of accurate, in-house AI classification is a game-changer. Sentra’s approach, backed by $50M in new funding, keeps data inside the security perimeter, slashing the exposure risk that comes with cloud-based large language models. This could redefine how organizations implement data loss prevention and zero-trust architectures.

The cybersecurity community is witnessing a paradigm shift in data classification, driven by the emergence of small language models (SLMs) that can operate entirely within customer environments. Sentra, a data security startup, through its CEO Yoav Regev, has announced that these advances now make it possible to classify massive volumes of unstructured data with high accuracy and efficiency, all while avoiding the privacy and cost pitfalls of large, cloud-based AI models. This development is poised to address one of the most persistent pain points in data security: the inability to effectively discover and label sensitive data at scale without external exposure.

“

Sentra’s approach, backed by $50M in new funding, keeps data inside the security perimeter, slashing the exposure risk that comes with cloud-based large language models.

Data classification has long been the Achilles’ heel of enterprise security. Organizations generate petabytes of unstructured data—emails, documents, chat logs, and more—that contain the majority of sensitive information but remain dark to traditional rule-based tools. Previous AI approaches tackled this by deploying large language models, but those required expensive GPU infrastructure and the exporting of data to external services, introducing latency, high costs, and serious compliance risks under regulations like GDPR and HIPAA. Regev, who led the cyber department in Israeli Military Intelligence for nearly 25 years, explained that customers rightfully expect accurate classification without such trade-offs. “The small models can give you these days amazing, amazing efficiency with high accuracy,” he said. “You can train them without the customer data, and you can run it inside.”

Sentra’s approach exemplifies a new wave of AI-powered data security posture management (DSPM). Its small models are trained without accessing actual customer data, thereby preserving confidentiality from the outset. Running the classification engine locally—whether on-premises or within the customer’s own virtual private cloud—eliminates the need to mirror data externally. This not only maintains data sovereignty but also shrinks the attack surface, as no additional copies of sensitive information are created outside the security perimeter. For security teams, that means fewer vectors for breach and lower auditing burdens. Regev emphasized that unstructured data represents “the biggest flow that you feed your models internally,” underscoring why mastering its classification is critical for any AI-driven enterprise.

The financial backing behind this vision is substantial. Sentra recently secured a $50 million funding round, as reported by GovInfoSecurity, signaling strong investor confidence in the cost-saving and security-enhancing potential of small AI models. This capital is earmarked to further develop its data protection platform, which threatens to disrupt legacy vendors like Varonis and BigID, who have historically relied on pattern matching and resource-heavy analytics. By commoditizing classification through efficient, local AI, Sentra effectively lowers the total cost of ownership for data security programs, a factor that becomes increasingly vital as economic pressures squeeze cybersecurity budgets.