Onyx Security Raises $113M Series B to Control Advanced AI Agents

Onyx Security, an AI control company, has raised a $113 million Series B led by Bessemer Venture Partners, with participation from Cyberstarts, TCV, Conviction, FirstMark, Vintage, QuantumLight, and G Squared.

The round brings Onyx’s total funding to $153 million since it was founded two years ago.

Revenue surge after stealth

Since emerging from stealth about four months ago, Onyx says it has grown revenue fourfold. In June, Anthropic announced an integration with Onyx to help secure enterprise AI adoption, on top of dozens of other AI providers already connected to the platform.

Onyx positions itself as a secure AI control plane: security, governance, and infrastructure teams can discover how advanced AI is used in their environment and intervene when agents go off-policy.

Using proprietary models, Onyx monitors steps in agent reasoning and can correct unintended or malicious actions at runtime. Customers include Fortune 500 firms across banking, technology, insurance, and energy. Revolut is among the named users.

The control problem

As more capable models land in production, employees are handing agents access to critical systems and complex jobs. That creates a new accountability gap: agents are empowered, but not always governed.

“In 2025, less than one percent of actions were made by agents. These new autonomous actors will soon be taking the overwhelming majority of actions, which begs the question, how do we guarantee these actions are legitimate?” Maxim Bar Kogan, CEO and co-founder, Onyx

Bessemer partner Amit Karp framed the opportunity as building the control layer enterprises need before they can adopt agents at scale.

Where the capital goes

Onyx says the new funding will train the next generation of its proprietary models and scale enterprise go-to-market in the United States and internationally.

Decoded Take

Onyx is selling a premise enterprises are starting to accept: agent adoption without a control plane is an operational risk, not just a security footnote. Runtime intervention matters more than policy PDFs, especially once agents touch money movement, customer data, and production systems. The Anthropic integration is a distribution signal; the test is whether Onyx becomes default infrastructure across model providers, or one more point tool in a crowded AI security stack.