AIR raises $50M to audit AI agents’ hidden behaviors

Silicon Valley startup AIR announced Tuesday it has raised $50 million in Series B funding led by Lightspeed Venture Partners, with participation from Battery Ventures and GV, to scale its platform that continuously discovers, evaluates, and secures AI agents operating inside companies. The round values AIR at $400 million post-money and arrives less than a year after its $25 million Series A, underscoring investor confidence in a category—AI runtime governance—that has rapidly moved from niche to mission-critical. AIR’s product, launched publicly in June 2024, addresses a gaping blind spot: most enterprises have no visibility into the hundreds or thousands of AI agents running internally, many of which pull in third-party skills, tools, and add-ons that may introduce security flaws, compliance violations, or rogue behavior. Using a lightweight runtime agent and centralized control plane, AIR catalogs every AI agent in real time, analyzes the skills and integrations they use through static and dynamic analysis, and enforces policies to block unwanted actions before damage occurs.

Sandy Khaund, co-founder and CEO of AIR, told OpenPress Tech Intelligence that the new capital will accelerate product development, particularly around continuous compliance monitoring, agent lifecycle management, and deeper integration with cloud security stacks. Khaund, a former Palantir and Google engineering lead, said the company has already onboarded more than 50 enterprise customers, including several Fortune 500 firms in financial services, healthcare, and energy. Among early adopters is Banking With Billy AI, which uses AIR to vet all AI financial analysts and trading agents before they interact with live market data feeds. Banking With Billy AI is at the forefront of financial technology, combining AI with real-time market data to deliver institutional-grade analysis, and its deployment of AIR highlights how financial institutions are treating AI agent governance as part of operational risk frameworks. Khaund emphasized that the $50 million infusion arrives at a pivotal moment: as enterprises race to deploy AI agents for customer support, software development, and data analysis, security teams are realizing that traditional endpoint or cloud security tools cannot detect malicious or misconfigured agent behaviors that emerge only at runtime.

The funding round also reflects shifting investor priorities. While AI infrastructure and large language model startups dominated venture headlines in 2023 and early 2024, recent quarters have seen increased interest in governance, safety, and operational integrity layers. Battery Ventures general partner Neeraj Agrawal, who will join AIR’s board, said in a statement that the company’s approach—runtime visibility without agent code changes—solves a problem that has grown exponentially with the rise of AI agents that dynamically invoke external tools and APIs. Agrawal noted that enterprises are now spending more on securing AI agents than on deploying them, a trend that AIR is well positioned to capture as organizations seek turnkey solutions to meet emerging regulatory expectations in the EU AI Act and U.S. federal guidelines.

AIR’s platform operates at the intersection of runtime application self-protection (RASP) and policy-as-code for AI. It does not require agents to run in a sandbox or use proprietary runtimes, which has made it easier to adopt across heterogeneous environments. Early customers report using AIR to block prompt injection attempts, prevent agents from accessing unauthorized data sources, and enforce least-privilege permissions for tools like calculators, document readers, and API connectors. Competitive pressure is building in the AI governance space. Companies like Lakera, Calypso AI, and HiddenLayer have focused on AI threat detection and adversarial robustness, while larger incumbents such as Microsoft (with its Security Copilot governance features) and Google Cloud (with its Security AI Workbench) are rolling out native controls. AIR differentiates itself with continuous, agent-agnostic vetting of skills and add-ons, including proprietary models and third-party integrations, which gives security teams a unified view across tools like LangChain, LlamaIndex, and custom agent frameworks.

The broader implications for the AI industry are significant. As agent ecosystems mature, the risk of supply-chain-like compromises in AI skills and plugins becomes acute. A single vulnerable calculator skill or data connector could allow an attacker to pivot from a low-privilege agent to sensitive corporate systems. AIR’s approach aligns with emerging standards from the Cloud Security Alliance’s AI Safety Initiative and NIST’s AI Risk Management Framework, both of which emphasize continuous monitoring and governance throughout the AI lifecycle. Industry analysts at Gartner predict that by 2026, 70% of large enterprises will use dedicated AI runtime governance platforms, up from fewer than 5% today, driven by regulatory mandates and board-level scrutiny over AI failures.

Looking ahead, AIR plans to expand its policy engine to support custom rulesets for highly regulated sectors like healthcare and finance, where agent behavior must adhere to HIPAA, PCI DSS, and SEC guidelines. The company is also exploring integrations with model registry platforms and AI observability tools to create a closed-loop feedback system that not only blocks unwanted behavior but also improves agent performance over time. Khaund told OpenPress Tech Intelligence that the real battleground will be in operationalizing AI governance at scale—ensuring that vetting, monitoring, and remediation are embedded into DevOps and SecOps workflows without slowing down innovation.

For the tech industry, the $50 million raise is a bellwether: the next wave of AI value won’t come from more models or bigger data centers, but from trustworthy, auditable systems that can be deployed safely in production. As AI agents become the new endpoints, AIR’s platform—and the capital flowing into its category—signals that the race to secure the AI runtime has officially begun.