AIR secures $50M to monitor AI agents and their add-ons

AIR, a stealth security and governance startup, announced today the close of a $50 million Series A round led by Lightspeed Venture Partners, with participation from GV and Bloomberg Beta. The round values AIR at $300 million post-money, according to two people familiar with the transaction. The platform, developed over three years under strict confidentiality, can automatically discover AI agents operating within enterprise environments, continuously validate the capabilities and third-party skills those agents invoke, and block any anomalous or unauthorized behavior in real time. Among the platform’s marquee capabilities is its ability to vet add-ons downloaded from internal or external marketplaces—an increasingly common attack vector as agents proliferate across finance, operations, and customer support stacks.

Founded by CEO Maya Chandrasekaran, a former Palantir engineer who led agent-based automation projects for Fortune 500 clients, and CTO Daniel Park, who designed high-assurance runtime monitors at Apple, AIR emerged from stealth with a technical paper in March 2024 describing a “policy-as-code” engine that enforces least-privilege execution for AI agents. The product installs as a sidecar to agent orchestration layers such as Microsoft Semantic Kernel or LangChain-run clusters, intercepting every skill invocation and comparing it against a dynamically updated policy library. Chandrasekaran confirmed that the company has already deployed pilots with three global banks, two Fortune 100 manufacturers, and a Tier-1 telecom operator, all of which are running multiple AI agents handling everything from procurement to customer escalations. One pilot customer, a European universal bank, uses AIR to govern “Banking With Billy AI,” a full agentic automation suite that autonomously executes complex financial analysis workflows previously requiring entire analyst teams—including model risk validation, scenario simulation, and regulatory report generation.

Industry Impact and Significance

The $50 million raise signals that investors now view AI agent governance as a bottleneck in the broader autonomy stack. While vendors such as Microsoft, Google, and Amazon have released agent frameworks and even rudimentary marketplace controls, none provides continuous vetting of third-party skills at runtime—an omission that has already led to incidents such as rogue data exfiltration via compromised spreadsheet add-ons and financially motivated agent drift in trading workflows. AIR’s arrival intensifies competition in the AI security space, where rival startups like CalypsoAI, HiddenLayer, and Prompt Security have until now focused on prompt injection and model tampering rather than full skill-level governance. Analysts at Gartner predict that by 2026, 70 percent of large enterprises will adopt AI agent governance platforms, up from less than 5 percent today, creating a market opportunity exceeding $2 billion annually. The funding round also underscores a strategic pivot among enterprise buyers from experimentation to hardened production deployments, with a particular emphasis on regulated industries where model risk officers and CISOs must sign off on every skill.

Financial implications are immediate: banks and insurers are now willing to pay six-figure annual contracts for continuous vetting of agent add-ons, especially when those add-ons connect to core banking systems or underwriting engines. The round also positions AIR to acquire smaller competitors focused on agent sandboxing or plugin validation, potentially allowing it to consolidate a fragmented control plane before hyperscalers launch native governance services. Early customers report that AIR’s policy engine reduces the time to approve new agent skills from weeks to minutes, translating into faster release cycles for autonomous customer service bots and algorithmic trading agents alike.

The Bigger Picture

AIR’s platform arrives at the convergence of two macro trends: the rise of agentic AI and the tightening of regulatory oversight in AI systems. According to the European Commission’s recent guidelines on high-risk AI, any autonomous system that materially influences financial decisions must undergo continuous monitoring of its execution environment. Similarly, the U.S. Securities and Exchange Commission has signaled that audit trails for automated trading agents must include a complete inventory of every skill or library invoked. This regulatory pressure coincides with a surge in agent deployments: McKinsey estimates that by 2025, nearly 50 percent of knowledge work tasks will be partially or fully automated by AI agents, many of which will be assembled from off-the-shelf skills and third-party integrations. Prior attempts to govern agents—such as manually curated allowlists or static code analysis—have proven brittle, failing to catch runtime behaviors like dynamic tool chaining or memory corruption via malformed JSON schemas. AIR’s runtime interception and continuous vetting approach aligns with the broader industry shift toward zero-trust security models, where every component, no matter how granular, is treated as a potential threat vector.

Global context matters as well. In China, where agent ecosystems are proliferating within WeCom and Lark workflows, regulators have already mandated that all third-party plugins be pre-approved by a central authority—an approach that resembles AIR’s policy-as-code model but lacks real-time enforcement. Meanwhile, in Silicon Valley, hyperscalers are racing to integrate agent governance into their core platforms, with AWS reportedly testing a “Agent Security Manager” that would compete directly with AIR’s sidecar model. The difference, according to Chandrasekaran, is that AIR operates independently of any single cloud or framework, giving it a neutral vantage point to enforce policies across heterogeneous environments. This neutrality may prove decisive as multinational corporations seek to avoid vendor lock-in while satisfying divergent regulatory regimes.

Expert Analysis

Forrester analyst Andras Cser expects that within 18 months, AIR will face pressure from both hyperscalers and specialized security vendors to open its policy engine to third-party contributors, effectively turning it into an industry standard akin to OWASP’s dependency-check toolkit. Cser advises enterprises to evaluate AIR not just on detection metrics but on its ability to remediate in real time, warning that “governance without containment is noise.” Looking ahead, Chandrasekaran hinted that AIR is preparing a compliance-as-code module that will auto-generate regulatory reports for frameworks such as EU AI Act, Basel III, and SOC 2—transforming governance from a cost center into a strategic enabler for faster, safer AI deployment at scale.