AIR Security Raises $50 Million to Build a Firewall for AI Agents

Key points

• AIR Security raised $50 million as it came out of stealth, meaning it moved from private development to public launch.
• The startup makes a firewall, a security gate that inspects traffic or instructions before they reach a system, designed specifically for AI agents.
• AIR's product screens AI skills, plugins and MCP servers for malicious instructions, excessive permissions and software supply chain risks.
• The product targets a gap that existing security tools were not built to cover: AI systems that act on behalf of humans.

A startup called AIR Security has stepped into public view with $50 million in funding. The question behind the product is one most companies haven't asked yet: who is checking what your AI tools are actually doing?

The company, first reported by SecurityWeek, makes what it calls an AI agent firewall. An AI agent is software that doesn't just answer questions but takes actions, booking meetings or placing orders on a user's behalf. A firewall is a security gate that sits between systems and decides what is and isn't allowed through. AIR's product applies that idea to AI agents.

What exactly does AIR's firewall look for?

Three things: hidden malicious instructions, permissions that are too broad, and risks buried in the software supply chain.

First, malicious instructions. AI agents can be fed commands through plugins, which are small add-on tools that extend what an AI can do, or through MCP servers. MCP, which stands for Model Context Protocol, is a technical standard that lets AI agents connect to external data sources and services. Criminals have found ways to hide harmful commands inside these channels, telling an AI to leak data or take unauthorised actions without the user realising. OWASP named "malicious skills" the number-one AI agent risk when it published its first security watchlist for AI agent skills on 21 August 2026.

Second, excessive permissions. An AI agent that can read your calendar probably doesn't also need access to your company's financial records. AIR's firewall flags situations where an agent has been granted far more access than its job requires, a classic setup for abuse.

Third, software supply chain risk. This refers to danger that enters through third-party code or services that an organisation didn't write itself but relies on. If a plugin was built by an unknown developer, or a connected service has been quietly altered, that risk flows into any AI agent using it.

Should organisations using AI tools be worried?

Yes, and not in a theoretical way. Businesses are connecting AI agents to real systems at speed. The security tooling has not kept pace.

Traditional firewalls watch for known attack patterns in network traffic. AI agents operate differently: they read natural language and make decisions. A firewall not designed for that context will miss a lot. Investors clearly see the problem as real: AIR's $50 million raise follows the $7.5 million that went to agent-control startup Xpander in August, which we reported at the time.

The practical step, if your organisation is already using AI agents: audit what permissions each agent holds and ask vendors what screening happens before third-party plugins connect to your systems.