Startup Emerges from Stealth with $50 Million for AI Security Enhancement
In a significant development within the technology sector, a startup known as Air, spearheaded by a former Israeli Military Intelligence leader, has recently launched with a robust funding of $50 million aimed at enhancing the interpretability of artificial intelligence (AI). Through this capital investment, the company aspires to enhance the security mechanisms surrounding AI agents, ensuring these systems remain safe while interacting with various online platforms and resources.
The financial support comes notably from prominent investors, including Sequoia Capital and Greenoaks. This funding enables Air to reverse-engineer AI models to allow organizations to assess their safety and reliability, as business processes increasingly incorporate AI technologies. Co-founder and CEO Yair Saban emphasized the urgent need for security teams to gain insights not just into the infrastructures that support AI models but also into the models themselves. Saban articulated a vision for the company that combines aggressive product development with expansive distribution strategies.
"Our primary aim is to develop a substantial research capability,” Saban declared to Information Security Media Group (ISMG). He conveyed that the entity that invests the most in research will ultimately dominate the field. Air, established in January 2026, currently comprises a team of 40 professionals and has made a considerable impact due to Saban’s extensive experience within Unit 8200 of the Israeli Military Intelligence, where he led research and development efforts for over eleven years.
Historically, Air’s research initiatives commenced with traditional cybersecurity challenges, focusing on identifying new risks associated with AI systems and combating online malicious campaigns. However, looking forward, Saban envisages a shift towards foundational AI research. The goal is to construct an elite research organization that addresses emerging security issues as dependency on AI grows among enterprises.
"Our current laboratory is primarily dedicated to cybersecurity risks,” Saban noted. He explained that developing interpretability within AI models — essentially reverse-engineering their operations — is crucial for ensuring their safety in a changing digital landscape, which he predicts will define industry challenges by 2027.
This heightened necessity for effective AI governance arises as many enterprises, which previously had little engagement with graphics processing unit (GPU) infrastructure, begin to download and run open-weight models internally. This transformation raises critical concerns regarding the origin of these models, their potential for alteration, and their functionality once integrated. Air aims to scrutinize these models pre-deployment, thereby allowing organizations to ensure that implemented models are genuinely safe.
"Consider the vast enterprises that had never utilized GPU technology, now downloading models from the internet and operating them on new GPU systems," Saban stated. "It is essential to guarantee that unauthorized actions such as cryptomining do not occur on their systems, and to ensure that no rogue models have been introduced."
Air’s strategic focus centers on imposing control over AI agents’ internet interactions. Saban predicts that these agents will increasingly facilitate enterprise operations on the web, challenging the conventional assumptions that have been historically applied regarding human browsing behavior.
"There’s a common refrain: ‘Don’t let your agents flounder around online,’" he remarked. For instance, he shared an anecdote about a company deploying an AI agent tasked with finding flight options, which fell victim to fraud because it lacked the capability to discern trustworthy websites.
Air’s methodology necessitates continuous evaluation of websites and other online resources, distinguishing it from entities that classify materials solely once. The company employs a four-stage vetting process: static analysis, dependency analysis, sandbox analysis, and continuous credibility assessment. The objective of this rigorous examination is to determine the trustworthiness of resources before an enterprise agent engages with them.
"I’m pleased to embrace a significant level of challenge. AI may simplify many processes, yet continuously vetting the myriad resources available on the internet is no easy feat," Saban admitted.
Air positions itself distinctly from AI security firms that prioritize runtime controls, as this framework creates a challenging balancing act. Security products must effectively obstruct malicious activities without compromising the operational integrity of agents or degrading the user experience. Instead, Air emphasizes “pre-runtime” security — a proactive approach akin to the vetting processes employed by app stores before users can install software.
"You don’t analyze applications at the moment of download — that’s far too late. The examination must occur during the upload to the Play Store,” Saban explained, drawing a parallel to their own vetting processes. Air’s prevention-focused model ultimately ascribes credibility to the resources that agents may encounter while operating online.
Organizations must exercise discernment over the parts of the internet their autonomous agents can traverse. This requirement becomes increasingly critical in light of sensitive data ranging from pharmaceutical records to financial information, where a breach could irreparably damage an organization’s credibility.
In an era where AI technologies are gaining unparalleled traction and prominence, Air’s innovative approach to safeguarding AI agents demonstrates an essential stride towards secure technological progression. The ambition behind Air resonates with a broader trend: the imperative need for enhanced security and reliability in an increasingly AI-integrated business landscape.