The trajectory of a breakthrough cybersecurity startup often begins long before there is a registered company, a formalized name, or a commercial product on the market. That was precisely the case with Yair Saban and Niv Hoffman. When venture capitalists first evaluated the founders, they ran extensive references across the competitive Israeli cybersecurity ecosystem, and the feedback was remarkably consistent: Saban and Hoffman were widely regarded as possessing the makings of one of the strongest founding teams in the country. One former commander of Unit 8200, the elite intelligence agency of the Israel Defense Forces, offered a striking endorsement of Saban, describing him as the only person the commander had ever encountered who possessed every single attribute required to one day command Unit 8200 itself—marking about as powerful an endorsement as exists within the high-stakes world of Israeli technology. Meanwhile, Hoffman emerged as a technical outlier who previously led Unit 8200’s most critical vulnerability research department, where he successfully cultivated a sprawling network of elite technical talent. The sheer velocity at which the team operates was underscored by a simple timeline: Saban finished his military service on a Thursday, and investors shook hands on a seed investment that very Saturday, setting the tone for the startup’s rapid ascent.
Long before Saban and Hoffman began architecting a commercial defensive security product, they engineered a proof-of-concept project known as MOAK, which stands for the "Mother of All KEVs." MOAK was designed as an agentic workflow capable of autonomously producing working exploits for freshly published vulnerabilities without any human intervention. In a striking public demonstration that caught the industry’s attention, MOAK successfully exploited a React vulnerability in a mere 21 minutes—a stark contrast to the estimated six days it typically takes a standard enterprise organization to roll out and apply a software patch. The founders publicly released the tool to sound an alarm across the cybersecurity community, warning that the technology sector was severely underestimating the immense leverage that malicious actors could rapidly extract from artificial intelligence, and emphasizing that defensive security tooling urgently needed to evolve to match this new threat landscape.
The underlying vulnerability stems from a structural shift in software development. The moment software developers began importing pre-packaged components and third-party libraries instead of writing every line of code from scratch, software acquired a complex supply chain. It subsequently took the technology industry two decades, alongside a series of painful and costly security incidents ranging from Log4j to SolarWinds, to construct the necessary machinery and protocols for inspecting incoming software components. Today, artificial intelligence agents face precisely the same systemic vulnerability, with one critical distinction: an agent’s supply chain is assembled dynamically at runtime, and in many enterprise environments, absolutely nobody is inspecting it. A modern enterprise agent regularly consumes skills, plugins, Model Context Protocol (MCP) servers, sub-agents, hooks, commands, and rules—each operating with the agent’s own high-level permissions, directly inside a user’s environment, and interacting with sensitive user data. In practice, these components are frequently installed with the casual carelessness that most individuals typically reserve for downloading an obscure browser extension.
To underscore the severity of the threat, Air spent several months systematically demonstrating the real-world stakes. The team authored a deliberately malicious AI skill and watched as multiple mainstream AI marketplaces routinely waved it through as entirely safe. Their published research, titled "The Story of Skills," documented how a researcher-built malicious skill successfully reached more than 26,000 active agents through a trusted marketplace and social media channels, with every automated security scanner completely clearing it of wrongdoing. Shortly thereafter, the team published "SkillJacking," a research initiative which revealed that 925 live skills were pulling their underlying dependencies and prompts from abandoned, dead links. Air proactively registered those dead links and effectively took control of approximately 134,000 corporate agents that had previously installed the compromised skills. Across all these tests, every public security scanner tested by the researchers missed the underlying vulnerabilities entirely.
Demonstrating their trademark methodology of identifying a critical flaw, proving its exploitability, publishing the findings transparently, and immediately shipping the tooling to close the gap, Saban and Hoffman released ScanAir on the same day as their research. ScanAir is an OWASP-aligned security scanner purpose-built for agentic skills. This consistent pattern of rapid, high-impact execution has become a defining characteristic of the young startup.
Today, Air offers a comprehensive suite of three distinct products designed to cover the entire lifecycle of an AI agent’s add-ons. The first product, Air Filter, vets every potential add-on before an agent is ever permitted to consume it, utilizing advanced static analysis, rigorous dependency checks, and secure sandbox detonation. It then actively intercepts unsafe installations in real time and offers safe alternatives in their place. The second product, Air Defend, monitors the actual runtime behavior of agents once they are active, providing enterprise security operations teams with robust detection and response capabilities tailored to agent activity. The final product, Air Control, manages overall posture and policy governance, enabling organizations to maintain visibility over which agents exist, what sensitive data or systems they are permitted to access, and allowing administrators to instantly revoke a malicious or compromised skill across thousands of agents simultaneously. The architecture can be understood through a home security analogy: Air Filter functions as the deadbolt locks and childproofing measures, Air Defend operates as the security cameras and the monitoring personnel watching the feeds, and Air Control serves as the master control panel by the front door where every entry point is accounted for, every access code is assigned, and all permissions can be revoked at once.
In the fast-paced ecosystem of Israeli enterprise technology, every few years the absolute highest-tier cyber talent originating from Unit 8200 consolidates around a select handful of promising early-stage teams. The wave began with companies like Wiz, Cyera, and Island, followed by subsequent cohorts featuring Eon, Glow, Decart, and Kela, and has now firmly converged around Air. Saban and Hoffman have demonstrated themselves to be sharp, deeply driven, and natural leaders who have proven their unique ability to attract extreme talent density, assembling a formidable team of top-tier engineers and researchers in just a matter of months. Sequoia Capital, which previously led the company’s seed funding round, is now doubling down on its investment by leading the startup’s Series A financing.
As modern enterprises increasingly move toward granting artificial intelligence agents genuine authority to execute actions across real-world systems and corporate databases, chief information security officers everywhere will inevitably face difficult questions regarding governance and safety: exactly what code is our agent running, where did those components originate, and who verified their security? Currently, very few organizations possess complete and definitive answers to these inquiries. However, with its comprehensive suite of security products, Air is positioning itself to build the definitive system of record for the age of enterprise AI agents.

Comments (0)