loading...
AIR discovers every AI agent running inside a company, vets their tools and permissions, and blocks unauthorized interactions — a security layer the agent era never had.
ENTRY ANGLES
Build compliance certification layer for regulated AI agent deployments (financial services, healthcare) · Create vertical-specific pre-vetted skill catalogs with use-case-specific vetting reports · Build cyber insurance underwriting integration — turn the audit record into an underwriting input
VERTICALS
CAPABILITIES
Enterprise security expertise, Agent ecosystem knowledge, Regulatory compliance understanding, Runtime monitoring infrastructure
A fake AI agent skill passed every existing security scanner and reportedly propagated to 26,000 agents before anyone caught it. AIR, a company that launched publicly two weeks ago, says 27% of agent skills available on the open internet fail its continuous vetting checks. Those two data points together explain why a $50 million seed round materialized in six months.
AI agent security has a structural blind spot: most enterprise security teams vet the agent, not the components the agent installs and calls at runtime. An agent deployed by IT gets reviewed and approved; the plugin it fetches three months later, from a third-party marketplace, does not. AIR is built around that oversight gap. Its platform has four layers: AIR Control discovers and inventories every agent running inside a company, including shadow deployments employees stood up without approval; AIR Filter vets skills, plugins, MCP servers, and sub-agents before installation; AIR Defend applies runtime enforcement — blocking untrusted tools, external instruction loading, and data exfiltration attempts as they execute; and a marketplace provides a catalog of pre-vetted, certified add-ons for enterprises that want to expand agent capability without building a vetting process themselves.
The founders, Yair Saban (CEO) and Niv Hoffman (CTO), are both veterans of Israel's Unit 8200 intelligence corps — the lineage that produced Wiz (Yinon Costica, an AIR investor), Orca Security, and Snyk. The company was founded in February 2026 and raised $10 million led by Sequoia Capital within its first weeks, followed by a $40 million round led by Greenoaks Capital. It emerged from stealth on September 1, 2026.
The comparison AIR keeps drawing is to software supply chain security — specifically to the Log4j moment, when a widely used dependency turned out to contain a critical vulnerability that affected thousands of applications whose teams had no idea they were using the component. Agent supply chains work on the same principle: an enterprise controls the agent it deploys but not the full graph of plugins, skills, and sub-agents that agent depends on at runtime. The agent is the application. The skills and MCP servers are the dependencies. Traditional endpoint security was not designed to inspect that layer.
What makes the current moment structurally significant is that agent skill ecosystems are growing before governance mechanisms exist for them. The major agent platforms are adding skill marketplaces faster than enterprises can evaluate what's in them. A company that deploys a hundred agents, each using four or five skills from third-party providers, has an agent supply chain with four or five hundred points of potential exposure — none of which appears on a conventional software inventory.
The 27% failure rate AIR cites across publicly available skills is not a statement about the sophistication of the attacks; it is a statement about the baseline quality of the supply. Most of those failed vets will be misconfiguration, excessive permissions, or unauthorized data access patterns rather than deliberate malice. The security risk is real in either case.
The immediate build that hasn't been attempted is agent compliance certification for regulated verticals. AIR's runtime blocking is a risk-reduction tool; what financial services firms and healthcare organizations additionally need is a documented audit trail — proof that at a given date and commit, every component used by a production agent had been reviewed against specific regulatory criteria. That's a compliance artifact, not a security product, and it reaches a different buyer (legal and compliance, not infosec) at a different price point.
The second opening is vertical-specific pre-vetted skill catalogs. AIR's marketplace is currently general-purpose, but vetting criteria differ fundamentally by use case: a healthcare agent needs to be checked against PHI access patterns and HIPAA constraints; a financial services agent against trading restriction rules and SOX reporting requirements. A compliance-first catalog organized by regulated vertical — where every listed skill ships with a use-case-specific vetting report — would be meaningfully differentiated from a general-purpose marketplace.
The longer-term bet is that AI coding tools, agent orchestration platforms, and enterprise software vendors will eventually need to partner with or acquire companies that own the verification layer. The company that establishes the vetting standard — whose approval becomes the de facto trust signal across agent marketplaces — is building something with the structural durability of a certification body rather than a point solution.