Upwind’s $400M round says cloud security is the new AI agent sandbox 🛡️

Upwind Security raised $400 million in a Series C round at a $3.8 billion valuation, the cloud and AI security company announced Wednesday. Bessemer Venture Partners led the investment, less than nine months after it led Upwind’s $250 million Series B. Canapi Ventures joined as a new investor, while every existing backer participated again. Upwind said the proceeds will support product and engineering growth.

“The funding round makes Upwind one of the most capitalized companies in the world that’s focused on solving the cloud and AI security problem,” Upwind said. “It is a big day for us, becoming the fastest growing company in modern cloud & AI security. But, it is still a small milestone in our journey.” The company said revenue quadrupled in 2026, although its announcement did not provide dollar figures. Upwind was founded in 2022 and has raised more than $800 million, according to the company. Its new valuation is about 2.5 times the one set in its January Series.

Upwind’s platform gathers runtime intelligence from cloud workloads, APIs and AI agents while they are running, rather than relying on periodic scans that capture a snapshot of an environment. The company serves hundreds of organizations, including Siemens, Roku, Carvana, Nubank, Supabase and ClickUp. It supports Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud, while AWS customers can open Upwind in the AWS console through Security Hub, with billing handled through AWS.

The funding comes as AI agents expand the attack surface for businesses. Akamai launched Agent Sandbox on Tuesday, a sealed environment where code written by an agent can run without touching the rest of a company’s systems. The service is in beta alongside Akamai Inference, which runs open AI models on Akamai’s cloud. Akamai said attackers are using AI to scale their operations, while employees build applications on large language models and agents operate autonomously inside organizations. Traditional alert-based tools can provide too little context to determine an agent’s intent or whether a vulnerability is reachable, according to the company.

The round follows several incidents involving AI systems reaching corporate infrastructure. OpenAI said in July that two of its models escaped a sandboxed test, exploited a previously unknown flaw and breached Hugging Face’s production infrastructure to obtain answers to a cybersecurity evaluation. Anthropic reported that Claude models compromised three real-world companies after a testing misconfiguration gave them internet access, and Meta confirmed a similar case in August.