Agentic AI
AI Security Startup AIR Lands $50M to Police Agent Tools and Skills
As autonomous AI agents gain wider access to corporate systems, a new platform promises to continuously monitor and vet the tools and add-ons agents rely on—spotting risky code before it runs.
TechCrunch — AI
Companies are increasingly entrusting AI agents with access to critical systems and databases, creating what amounts to a new software supply chain. That opens the door to attacks: instead of targeting an agent directly, adversaries can poison the tools, plugins, and external data sources the agent consumes. A new Israeli security startup believes enterprises will need independent oversight to manage this emerging risk.
AIR, founded by cybersecurity veterans Yair Saban and Niv Hoffman, has emerged from stealth with $50 million in fresh capital to tackle the problem. The company claims it can discover all AI agents running inside an organization, continuously inspect the skills and add-ons they load, and block access to tools that fail security checks.
The Supply Chain Problem#
The startup frames its pitch around an apt parallel. In the early 2000s, software drivers loaded into the OS kernel without any verification. Today, every driver must be cryptographically signed, preventing malicious code from being accidentally installed. AI agent plugins and skills, Saban argues, face no such gatekeeping—despite being equally capable of executing code within enterprise infrastructure.
AIR's approach mirrors kernel-level security controls: it maintains a continuously updated whitelist of vetted add-ons by scanning the wider ecosystem for changes and malicious behavior. The platform already filters out roughly 27% of publicly available skills it encounters, suggesting a real problem in the wild.
How AIR Works#
The startup's platform operates on three layers:
- Discovery: Finding all agents active across a company's environment, including those using unapproved tools or personal accounts
- Enforcement: Intercepting and analyzing actions in real time, such as when an agent attempts to load a skill or fetch content from the internet
- Curation: Cross-referencing requested tools against AIR's maintained whitelist of vetted components
Saban notes that the hard part is not initial scanning but continuous re-verification—a skill approved last month may become dangerous if its developer's account is compromised or upstream dependencies change.
Market and Competition#
AIR claims over 20 enterprise customers, with roughly a quarter being large organizations. The strongest demand has emerged in regulated sectors including financial services and pharmaceuticals. Yet the company faces competition from other agent-security platforms: , , , and all offer overlapping discovery, control, and monitoring capabilities. Recent funding in the category—including Zenity's $125 million Series C in August—signals substantial investor confidence in the space.

