Onyx Security raises $113 million for AI agent control plane

Why it matters

AI agents can act with corporate credentials at machine speed. Onyx Security's round shows investors expect runtime control of those actions to become a large security category.

Maxim Bar Kogan and Gil Elbaz raised $113 million for Onyx Security on July 29, giving the founders fresh capital to build a control layer for the growing number of AI agents acting inside corporate systems.

Bessemer Venture Partners led the Series B, with Cyberstarts, TCV, Conviction, FirstMark, Vintage Investment Partners, QuantumLight and G Squared participating, according to Onyx Security's announcement. CTech reported an estimated $640 million valuation. Onyx Security did not state a valuation in its own announcement.

The financing brings Onyx Security's disclosed funding to $153 million. It follows a $5 million seed round and a $35 million Series A led by Conviction, which the founders disclosed when Onyx Security emerged from stealth on March 12, 2026. Onyx Security was founded in 2024, making the nine-figure round roughly four months after its public launch an unusually compressed financing schedule.

The evidence Onyx Security has chosen to publish is weighted toward product activity rather than conventional business metrics. Bar Kogan says Onyx Security secures more than 1.1 million agents and inspects more than 66 million AI sessions in real time. The founders have not supplied revenue, annual recurring revenue, pricing, customer count or growth figures. At its March launch, Onyx Security said its control plane was already being used by Fortune 500 companies, though it did not identify them.

A founder thesis built around losing control

Bar Kogan began working on Onyx Security after ChatGPT's release left him preoccupied with a specific question, according to Onyx Security's Series B post: "when something smarter than you is responsible for the things you depend on, how do you stay in control?"

His background combines engineering, product leadership and offensive cybersecurity. Bar Kogan served in Unit 8200, studied at the Open University of Israel and later became vice president of product and engineering at Mixtiles. He also won Israel's CodeGuru competition in 2008, according to his biography on Onyx Security's site.

Elbaz brought the AI side of the founding pair. He previously served in the Israeli Air Force's Operational Technology Unit and co-founded Datagen, which built synthetic data for computer-vision systems. Before starting Onyx Security, he worked as an AI architect in Nvidia's CTO office on infrastructure for training large language models.

The pair connected through running and spent their time discussing the consequences of letting agents operate critical infrastructure, Bar Kogan wrote in Onyx Security's launch post. Those conversations shaped Onyx Security's bet on supervisory software that watches other AI systems.

Their timing rests on a change in what enterprise AI can do. Chatbots mainly returned information to a user. Agents can receive credentials, call software tools, alter databases, write and deploy code, move files, contact customers and trigger workflows. Existing identity and security products can observe parts of that activity, but those products were generally designed around deterministic software and human users.

Onyx Security is betting that agents require their own enforcement layer.

Onyx Security wants to intervene before an agent acts

The founders describe the Onyx Security platform as a secure AI control plane. It discovers agents, models, AI applications and connected tools across browsers, endpoints, software services and cloud environments. It then inspects an agent's proposed actions against corporate policies before those actions reach another system.

Onyx Security's Guardian Agent can block an action, require human approval, reduce an agent's permissions or steer it toward a safer alternative. That runtime intervention is the core product bet. Security teams have traditionally investigated alerts after software or a user performed an action. An agent working at machine speed can complete a chain of consequential actions before a human analyst has time to examine the first alert.

The technical challenge is substantial. The control system must interpret an agent's intent, identity, context and requested action quickly enough to avoid becoming a bottleneck. It also has to distinguish malicious behavior from legitimate workflows that vary from one execution to the next. False positives that repeatedly stop useful agents would turn the security product into a tax on AI adoption.

Bar Kogan is using the new capital to train the next generation of Onyx Security's models and extend sales, business development and marketing in the United States and other markets. Onyx Security lists offices in Tel Aviv and New York.

Investors are funding a category before it settles

Onyx Security is entering a market with several competing definitions. Vendors are selling AI asset discovery, model testing, prompt-injection defense, data controls, agent identity, governance and runtime intervention, often under the same broad AI-security label.

Capital is arriving quickly. Hush Security raised a $30 million Series A on July 28 for AI-agent governance. Willow launched with $7 million in June to build identity and access controls for agents. In August 2025, SentinelOne agreed to acquire Prompt Security, showing that established cybersecurity companies were already buying their way into the category.

Bessemer's investment places Onyx Security among the best-capitalized independent contenders while enterprises are still deciding which controls should live inside existing security products and which warrant a separate platform. Onyx Security's board now includes Bessemer partners Amit Karp and Sameer Dholakia, alongside representatives from Cyberstarts and Conviction and the co-founders of Armis.

Bar Kogan is using the new capital for product development and go-to-market expansion. The $640 million estimate prices in a considerable share of that ambition only four months after the public launch. Building a durable business from that valuation will require converting large volumes of inspected agent activity into repeatable enterprise contracts, while proving Onyx Security can intervene without slowing the systems customers bought it to protect.