CrowdStrike Just Defined AI Agents as the New Insider Threat. The Agentic SOC Is Not a Product Feature — It's a New Enterprise Security Category.
With 17,800+ public AI add-ons catalogued and $435M flooding into agent security in five months, AIR's inline firewall is the first product designed specifically to intercept supply chain attacks before they reach agent context.
AIR Security emerged from stealth on September 1, 2026, with $50 million in seed funding to build an inline firewall for AI agents. Sequoia Capital led an initial $10 million round. Greenoaks Capital Partners followed within weeks with a $40 million investment. The company's founders, Yair Saban and Niv Hoffman, are veterans of Unit 8200, Israel's signals intelligence unit, with careers spent operating in the offensive security domain they are now defending against.
The timing was not incidental. AIR's launch research catalogued 17,800 publicly available AI add-ons with 6.7 million installs — MCP servers, agent plugins, skills, and tool integrations that give AI agents their capability by connecting them to external systems, data stores, and actions. These add-ons are also the attack surface that no enterprise security tool was designed to cover. When an AI agent calls an MCP server, retrieves a tool plugin, or invokes a third-party skill, it is trusting an external component with authority to influence its next action. AIR's firewall screens that trust relationship before the agent acts on it.
The $435 million that investors deployed across 12 AI agent security financings in the five months before AIR's launch signals that the trust relationship has been recognized as a category-level problem. The category barely existed as a named segment twelve months ago. Now it is absorbing institutional capital at seed stage because every organization deploying agents at scale is confronting the same exposure: the add-ons their agents trust are unverified, continuously updated, and largely invisible to existing security tooling.
The Attack Surface No One Was Measuring
The enterprise AI security discussion of 2025 and early 2026 concentrated on two threat vectors: prompt injection attacks against user-facing AI systems and data governance for AI training pipelines. Both are real. Neither addresses the supply chain.
Supply chain risk in software has been understood since at least the 2020 SolarWinds attack, which compromised 18,000 organizations through a single vendor's update mechanism that most customers had no process for auditing. The analogy in AI agents is structural. An enterprise deploying an AI agent that calls a third-party MCP server is trusting that server with real-time influence over the agent's reasoning and actions. If that server is compromised, misconfigured, or maliciously crafted, it can instruct the agent to take actions the enterprise would never permit if it reviewed the instructions directly.
The difference from traditional supply chain attacks is speed and invisibility. A compromised software library takes effect the next time a developer ships code and that code runs in production — there is a code review gate, a deployment pipeline, potentially a security scan, and a build artifact that can be inspected. A compromised MCP server takes effect the next time an agent calls it: in real time, dozens of times per session, without any equivalent review gate. The add-on update that introduced the malicious payload was pushed by the publisher; the enterprise has no visibility into it until the agent has already acted on it.
AIR's catalogue of 17,800 public add-ons includes entries from categories that most enterprise security teams would struggle to audit manually: coding assistants, search tools, database connectors, file managers, communication platform integrations, and customer data APIs. Many were published by individuals rather than verified organizations. Some had not been updated since their initial release. Others had transferred publisher accounts in ways that suggested supply chain acquisitions — the same attack pattern that compromised the event-stream npm package in 2018, now replicated in the AI add-on ecosystem.
The AI training opt-out coalition among Google, Microsoft, OpenAI, and Anthropic addresses what publishers put into training data. The AIR problem is different: it is about what agents pull from the web in real time, during inference, as live tool calls that execute with the agent's full authority.
How AIR's Agent Firewall Works
AIR's firewall operates inline between an AI agent and the add-ons it wants to call. The architecture has three enforcement modes that can be deployed independently or in combination:
Discovery mode catalogs every add-on, plugin, MCP server, and skill in an organization's AI environment — including ones the security team did not know were deployed. AIR's research found that most enterprise AI deployments have a significant shadow add-on problem: employees connect third-party tools to their AI assistants without IT review, mirroring the shadow IT patterns that created the cloud application governance crisis a decade ago. Discovery mode builds the inventory without blocking anything; it is the prerequisite for all other enforcement.
Evaluation mode runs each discovered component through a multi-layer risk assessment. Static analysis inspects the add-on's code and configuration for known malicious patterns, excessive permission requests, and supply chain provenance indicators. Dynamic analysis runs the add-on in a sandboxed environment to observe its actual behavior — whether it makes network calls it didn't disclose, accesses data beyond its stated scope, or sends outbound requests to unexpected destinations. Provenance analysis checks the add-on's author, update history, and permission scope changes over time. An MCP server that claims to be a document summarizer but requests API access to the organization's CRM is flagged for excessive permissions, the same principle app store review applied at the add-on invocation layer.
Enforcement mode blocks or quarantines flagged add-ons before an agent can act on them, generates audit logs for every add-on invocation, and routes high-risk decisions to human review queues. The enforcement layer deploys through existing MDM or EDR tooling rather than requiring new agent infrastructure — a deployment choice that reflects how security tooling actually gets adopted inside large enterprise environments.
Sequoia's Bogomil Balkansky, who led the firm's investment, identified the core design requirement: continuous re-verification rather than point-in-time approval. An add-on that was clean in January may be compromised in March. The components an agent uses change continuously as publishers push updates. A static approval-and-forget workflow misses the ongoing drift between what was approved and what is now running.
Why MCP Servers Are the New Attack Vector
The Model Context Protocol has become the de facto standard for how AI agents connect to external tools. It was designed to solve a real problem — standardizing the interface between AI agents and the capabilities they need to function — and it succeeded well enough that thousands of MCP servers are now publicly available. The ecosystem grew faster than the security infrastructure around it.
The attack potential of a compromised MCP server is higher than a compromised traditional API for a structural reason: AI agents treat the instructions they receive from tool calls as part of their reasoning context. A compromised REST API endpoint returns bad data; the developer or monitoring system notices and investigates. A compromised MCP server can return instructions that an AI agent will interpret as legitimate guidance for its next action. The agent has no intrinsic mechanism for distinguishing trusted instructions from adversarially injected ones — it processes both as context.
Prompt injection via MCP server responses was documented as an attack vector by mid-2026. A researcher at Trail of Bits demonstrated that a maliciously crafted MCP server response could cause a commercial AI coding assistant to exfiltrate source code to an external endpoint while presenting a normal-looking response to the user. The attack required no credentials, exploited no software vulnerability, and left no log entry in the application layer. It worked entirely by abusing the agent's default behavior of treating all context-window input as valid.
AIR's supply chain screening addresses this attack vector at the authorization layer rather than the content layer. By evaluating what an MCP server is allowed to say before an agent can invoke it, AIR intercepts potential attacks before the agent has processed any potentially malicious content. This is structurally more robust than content-based filtering — scanning what the MCP server returns — because a sufficiently sophisticated attacker can craft instructions that bypass content filters while still redirecting agent behavior. Authorization-layer controls block the conversation before it starts.
The Competitive Landscape: $435 Million in Five Months
The agent security category formed, reached critical mass, and attracted major capital in a compressed window. The vendors now competing in it are addressing different layers of the same underlying problem.
| Company | Focus Area | Key Investors | Funding |
|---|---|---|---|
| AIR Security | Add-on supply chain firewall | Sequoia, Greenoaks | $50M seed |
| Noma Security | AI governance and risk management | Unusual Ventures | $25M+ |
| Zenity | Agentic workflow risk | Bessemer, Vertex Ventures | $38M |
| Astrix Security | NHI and service account security | Bessemer, CRV | $45M |
| Operant AI | Runtime agent behavioral analysis | Sorenson, True Ventures | $20M |
| Sentra | Agentic data governance | Capital One Ventures | $150M |
Sentra's $150 million round addresses data governance for AI agents — what data agents can reach in enterprise data stores and whether that access is appropriate. That is adjacent to but distinct from AIR's supply chain screening, which addresses what external components agents can trust rather than what internal data agents can access.
CrowdStrike's agentic SOC capability, announced at Fal.Con Europe in October 2026, uses behavioral baseline detection to flag anomalous agent behavior after the agent has been running in production. That is a detective approach. AIR's firewall is a preventive approach. By the time behavioral detection catches an anomaly, a compromised add-on has already influenced the agent's reasoning.
Palo Alto Networks' AI agent security features in Prisma Cloud focus on detecting prompt injection attacks against user-facing AI systems. That coverage addresses a different threat vector — attackers manipulating agents through user-submitted content — rather than the add-on supply chain.
The practical implication for enterprise security buyers: no single vendor in this table covers the full agent security stack. The layers are genuinely different; the technical requirements of each are distinct; and the coverage questions — do you have visibility into what add-ons your agents use, and are you detecting anomalous behavior once they're deployed — require different tools to answer.
The Sequoia Thesis: Why This Round Was Structured as a $50 Million Seed
A $50 million seed round is unusual. Enterprise security seed rounds typically range from $5 million to $20 million, with Series A rounds used to fund the transition from early customers to scaled go-to-market. The AIR structure — $10 million from Sequoia, $40 million follow-on from Greenoaks within weeks of the initial close — reflects a specific thesis about category timing.
Sequoia's argument, as conveyed through Bogomil Balkansky's public statements, is that agent security supply chain tools will become infrastructure — something every enterprise deploying agents at scale will require as a matter of operational necessity, not as an optional security enhancement. Infrastructure tools in security have historically been won by the vendor that establishes market presence before enterprise security procurement cycles solidify. When an enterprise security team has selected a supply chain governance tool and integrated it into their agent deployment workflow, the switching cost is high enough that displacing the incumbent vendor requires a significant capability gap.
The window for establishing that market presence is narrow. Enterprise security teams are making agent deployment decisions now. The design partnership contracts and production deployments that AIR is funded to close in the coming quarters will determine whether it enters those procurement cycles as an established product or as an evaluated alternative to an incumbent.
The $50 million seed also prices AIR as a potential category leader at inception, not as a Series A company with proven revenue. For enterprise security buyers, that pricing is itself informative: when two of the most rigorous firms in venture capital structure a $50 million seed for a company in its first days of public existence, they are signaling their conviction that the problem is both real and large enough to support a standalone platform business.
Five Steps for Enterprise Security Teams Before Deploying Agents at Scale
The AIR launch and the broader $435 million flowing into agent security is a forcing function for enterprise security teams that have been deploying AI agents under conventional application security frameworks. Those frameworks were built for humans using software, not for software making decisions with agentic authority. The gap requires specific action.
1. Build a complete inventory of AI add-ons in your environment. Most enterprise security teams cannot currently answer the question: what MCP servers, plugins, and skills are deployed against our AI systems, and by whom? Shadow add-on deployment mirrors the shadow IT patterns of the early 2010s, when employees connected personal cloud storage and SaaS tools before IT policies caught up. Build the inventory now. Enumerate every add-on accessible from any corporate AI deployment, including employee-initiated deployments. The discovery is prerequisite to everything else.
2. Classify each add-on by risk tier before granting production access. Not all add-ons carry equal risk. An MCP server that connects an AI agent to an internal documentation system carries different risk from one that connects it to external payment APIs or customer PII stores. A risk framework should consider the data the add-on can access, the external connections it can initiate, the permissions it requests relative to its stated purpose, and the provenance and update history of its publisher.
3. Establish a formal approval workflow for new add-on deployments. The pattern in which employees connect third-party tools to AI systems without security review must be replaced with a lightweight approval workflow that creates a record. The workflow does not need to be burdensome — form-and-review by a security engineer is sufficient for most add-ons. What matters is that approval creates a record and creates friction for deployments that would not survive security scrutiny.
4. Implement continuous re-verification, not point-in-time approval. An add-on approved in January may push a malicious update in October. Add-on approval frameworks that treat approval as permanent create a false sense of security. Monitoring for approved add-ons must include automated alerts for version changes, permission scope changes, and publisher changes that could indicate a supply chain compromise.
5. Define incident response playbooks for agent supply chain events. When a compromised add-on is discovered in a production AI system, the forensic questions differ from conventional endpoint compromise: Which agents were connected to this add-on? What actions did those agents take while connected? What data did the add-on have access to? What instructions might it have injected into the agents' context? These questions need pre-answered process before the first real incident occurs.
Where AIR Fits as the Agentic Enterprise Matures
The agent security market's end state is genuinely uncertain. The category is nine months old. The major platforms — CrowdStrike, Palo Alto Networks, Microsoft — are adding agent security features that will cover portions of the problem. Some of the standalone vendors will be acquired. Some will find durable niches. A few will reach category leadership.
AIR's bet is that supply chain trust — specifically, verifying what add-ons can say and do before an agent trusts them — is a problem that behavioral detection platforms will not fully solve, because it operates at the authorization layer rather than the detection layer. The enterprise security teams that will close the full agent security gap will need preventive supply chain screening in addition to the detective behavioral analysis that platform vendors are building.
Whether that gap is large enough to support a standalone platform company or whether it gets absorbed into platform contracts within two years is the question AIR's $50 million is meant to answer. The company launched a week ahead of CrowdStrike's behavioral detection announcement, in the same week that Sentra closed its $150 million round. The category is real, the capital is committing, and the enterprise deployment decisions that will determine market structure are happening now.
Takeaway: AIR Security's $50 million seed from Sequoia and Greenoaks addresses a supply chain attack vector in AI agents that existing enterprise security tools are not built to handle. The 17,800 public AI add-ons catalogued at launch, with 6.7 million installs, represent an active and largely unmonitored attack surface that grows with every new MCP server, plugin, and skill added to the ecosystem. Enterprise security teams deploying AI agents at scale need visibility into what add-ons their agents use, a risk classification framework for those add-ons, an approval workflow that prevents shadow deployment, continuous monitoring for add-on changes, and incident response playbooks that address the specific forensic questions a compromised add-on raises. AIR's firewall approach is one answer; a disciplined internal governance program built on AIR's five-step framework is another; in practice, production-scale agent deployments at large enterprises will need both.
Frequently Asked Questions
What is AIR Security and what does its AI agent firewall do?
AIR Security is an enterprise security company that emerged from stealth in September 2026 with a firewall designed specifically for AI agents. Unlike traditional firewalls that inspect network traffic or endpoint firewalls that inspect processes, AIR's firewall sits between an AI agent and the third-party add-ons it wants to call — MCP servers, plugins, skills, and tool integrations. When an agent attempts to invoke an external add-on, AIR intercepts that call, evaluates the add-on against its risk database, and either permits the call, blocks it, or routes it to a human review queue. The evaluation covers static code analysis for known malicious patterns, dynamic sandbox testing of the add-on's behavior, and supply chain provenance checks that verify the add-on's author, update history, and permission scope. The goal is to prevent compromised or maliciously crafted add-ons from influencing agent reasoning before the agent has acted on potentially malicious instructions.
How much did AIR Security raise and who invested?
AIR Security raised $50 million across two closely timed seed rounds. Sequoia Capital led an initial $10 million seed. Greenoaks Capital Partners followed with a $40 million investment, closing within weeks of the Sequoia round. Additional investors include Swish Ventures and Netz. Sequoia's investment was led by partner Bogomil Balkansky, who articulated the investment thesis as continuous re-verification: the problem with AI agent add-ons is not point-in-time scanning but the need to continuously re-verify components that change underneath the agent as publishers push updates. The $50 million seed total is notably large for a company at inception stage; it reflects investor conviction that the agent supply chain security category will become infrastructure-grade for enterprise AI deployments.
What is an AI agent supply chain attack and how does it work?
An AI agent supply chain attack exploits the trust relationship between an AI agent and the third-party add-ons it uses. When an agent calls an MCP server, retrieves a plugin, or invokes a skill, it treats the response as part of its reasoning context. A compromised MCP server can return instructions that the agent interprets as legitimate guidance — not malicious content, but content that redirects the agent's actions toward attacker-controlled outcomes. This is structurally distinct from traditional supply chain attacks like SolarWinds, which compromised software binaries shipped in legitimate update packages. Agent supply chain attacks occur at inference time, in real-time tool calls, without any code review gate. A researcher at Trail of Bits demonstrated in June 2026 that a maliciously crafted MCP server response could cause a commercial AI coding assistant to exfiltrate code to an external endpoint while presenting a normal response to the user. The attack required no credentials and exploited no software vulnerability.
How is AIR Security different from CrowdStrike or Palo Alto Networks for AI agent security?
AIR Security and the platform security vendors address different layers of the agent security problem using fundamentally different approaches. CrowdStrike's agentic SOC capability, announced at Fal.Con Europe in October 2026, uses behavioral baseline detection: it builds a fingerprint for each deployed agent and flags behavioral deviations after the agent has been running in production. Palo Alto Networks' AI agent security features in Prisma Cloud focus on detecting prompt injection attacks against user-facing AI systems. AIR's firewall operates at the authorization layer, before an agent has acted on any external input: it evaluates what an add-on can say and do before the agent is permitted to invoke it. The key difference is preventive versus detective. By the time behavioral detection flags an anomaly, a compromised add-on has already influenced the agent's reasoning. AIR intercepts at the trust decision point. In practice, enterprise security teams deploying agents at scale will likely need both: AIR-style supply chain screening to prevent compromise, and CrowdStrike-style behavioral detection to identify compromises that the supply chain screen missed.
What are MCP servers and why are they a security risk for AI agents?
MCP servers implement the Model Context Protocol, a standardized interface that allows AI agents to call external tools, retrieve data, and take actions in connected systems. MCP was designed to solve a real problem — giving AI agents a consistent way to connect to capabilities outside their base model — and it succeeded: by mid-2026, thousands of MCP servers were publicly available, covering document management, database access, search, communication platforms, payment systems, and more. The security risk is structural. AI agents treat the information and instructions they receive from MCP server responses as part of their reasoning context. A maliciously crafted MCP server can inject instructions that redirect agent behavior toward attacker-controlled outcomes, without exploiting any software vulnerability. Additionally, MCP servers published by third parties may have excessive permissions, poor update hygiene, or supply chain compromises that make them vectors for data exfiltration or unauthorized access. AIR's catalogue found 17,800 public AI add-ons — many of which are MCP servers — with 6.7 million installs, representing an active and largely unmonitored attack surface.
What steps should enterprises take to protect against AI agent supply chain attacks?
Enterprise security teams deploying AI agents at scale should take five actions in sequence. First, build a complete inventory of every MCP server, plugin, and skill deployed against any corporate AI system, including employee-initiated deployments that security teams may not be aware of. Second, classify each add-on by risk tier based on the data it can access, the external connections it can initiate, and the permissions it requests relative to its stated purpose. Third, establish a formal approval workflow for new add-on deployments that creates a record and prevents shadow deployment. Fourth, implement continuous re-verification monitoring for approved add-ons, including automated alerts for version changes, permission scope changes, and publisher changes that could indicate a supply chain compromise. Fifth, define incident response playbooks specifically for agent supply chain events — the forensic questions differ from conventional endpoint compromise and must be pre-answered. Organizations that cannot yet deploy a commercial tool like AIR's firewall should address these five areas through internal governance processes while evaluating supply chain security tooling.