Context firewall that vets every input an agent takes in, from add-ons to web content to internal data

AIR Security was founded in February 2026 by chief executive Yair Saban and chief technology officer Niv Hoffman, and emerged from stealth in September 2026 with $50 million raised across two rounds, a $10 million round led by Sequoia Capital followed by $40 million led by Greenoaks. The company's premise is that the agent, not the model, is the exposed surface, and that the untrusted material reaching it arrives on three paths: the add-on ecosystem of skills, plugins, MCP servers, and sub-agents, the external web content an agent reads, and internal data such as emails and documents. Its published research found more than 17,800 public AI add-ons, representing about 6.7 million installations, that depend on instruction sources the installer does not control.

The product works in two phases. Before deployment it scans an add-on's manifest, code, permissions, and remote instruction sources for hidden behaviors, over-broad scopes, and supply chain tampering, and because a package that passed review can be altered upstream afterward, that vetting continues for the life of the installation. At runtime it sits on all three input paths and filters content before it enters the agent's context, catching prompt injection, hijacking attempts, permission abuse, and exfiltration paths, and blocking the resulting action. Policy is applied across the fleet to agent configuration, identity, and permissions, covering agents the organization sanctioned and ones it discovers. The company reports finding AI skills in circulation that impersonated Anthropic and OpenAI to pass security review and execute arbitrary code.