Onyx Security’s $113M Series B Puts Human Control at the Center of Enterprise AI Agents

Onyx Snags $113M to Keep AI Agents on a Human Leash

Onyx Security, an Israeli startup that emerged from stealth only four months ago, has closed a $113 million Series B round led by Bessemer Venture Partners. The deal values the company at around $640 million, according to Calcalist, and brings its total funding to $153 million. The rapid raise reflects a booming market for tools that govern how AI agents interact with critical business systems.

Onyx sells what it calls a secure AI control plane. The software sits between an enterprise’s AI agents and the systems they touch, mapping every agent, tracing its reasoning, and inspecting each action in real time. Any step deemed risky is blocked before execution. The platform now secures more than 1.1 million agents and inspects over 66 million AI sessions, the company says, with customers that include Fortune 500 firms in banking, energy, healthcare, and insurance—neobank Revolut is among them.

Chief executive Maxim Bar Kogan frames the problem as an impending flip in who does the work. In 2025, less than 1% of enterprise actions were taken by AI agents, he notes, but soon the overwhelming majority will be. “How do we guarantee these actions are legitimate?” he asked. Onyx’s answer is a runtime shield that stops an agent’s bad decision from having consequences in the physical or digital world.

The new capital will be used to train Onyx’s own models and to expand in the United States, the company said. With more than 80 staff in Israel, the U.S. and Canada, Onyx is positioning itself as the leash on a coming stampede of autonomous actors.

A New Front in Enterprise Security as Agents Scale

The Accountability Gap Driving Onyx’s Growth

AI agents are being handed credentials and access to critical systems without the governance that has always surrounded human employees. Onyx’s product fills that void by acting as an always-on auditor and enforcer. In sectors like energy, finance, and healthcare, where a machine-speed error can cause an outage, a market event, or put a patient at risk, the cost of an unblocked action can be enormous. Enterprises are increasingly unwilling to deploy agents without this layer, which explains why Onyx says its revenue has quadrupled since exiting stealth.

A Crowded Race to Secure the Autonomous Enterprise

Onyx is far from alone. Securing AI agents has become one of cybersecurity’s hottest corners, packed with Israeli rivals and already producing a billion-dollar acquisition. The investment case is that agent governance will become a must-have, not a nice-to-have, creating very large companies in the process. Onyx’s edge, according to its backers, is demonstrated traction: $113 million and a $640 million valuation four months after stealth, with a customer list that spans heavily regulated industries.

What the Anthropic Deal Signals

In June, Anthropic integrated Onyx to help its enterprise customers adopt AI safely. This partnership is a signal that even the builders of frontier models see a need for independent control layers. It also suggests that enterprises using multiple models will want a model-agnostic security tool, not one tied to a single AI provider. Onyx’s founders, veterans of Israeli intelligence and the Air Force, are betting that the harder problem is not building the agents but keeping them on a leash—and the market is starting to agree.

Practical Steps for Enterprises Navigating the Agentic Wave

For security and IT leaders already deploying AI agents, or planning to, Onyx’s rapid rise and the scale it is already handling point to several concrete steps:

• Map your agent surface immediately. Onyx’s claim of securing 1.1 million agents shows many enterprises already have a large and growing population of autonomous actors. Run an inventory; if you don’t know how many agents you have and what they are touching, you cannot secure them.
• Evaluate runtime intervention tools that inspect reasoning, not just outputs. Static permissioning cannot cope with agents that chain decisions. The ability to watch an agent’s reasoning chain and block a harmful action at runtime—exactly what Onyx does—is becoming the new baseline.
• Plan for model-agnostic controls. Anthropic’s integration of Onyx is a reminder that agent governance should work across AI providers, not just one. When choosing a platform, prioritize independence and broad model support over point solutions tied to a single LLM.
• Expect board-level attention soon. As agents move from handling less than 1% of enterprise actions to a majority, the risk profile changes. Security teams should prepare to brief the board on agent governance now, before an incident forces the conversation.