Onyx Security's $113M Series B Values AI Agent Control at $640M

The fastest-growing line item in enterprise security budgets may now be the software that watches the other software. Onyx Security, a two-year-old company selling what it calls a Secure AI Control Plane, has closed a $113 million Series B led by Bessemer Venture Partners, with Cyberstarts, TCV, Conviction, FirstMark, Vintage, QuantumLight and G Squared participating1620. The round, announced July 29, lifts total funding to $153 million and, according to press reports, values the company at roughly $640 million2811.

The valuation is not the most striking part. The clock is. Onyx emerged from stealth roughly four months before the announcement, having previously raised $40 million across a $5 million seed and a $35 million Series A, and says it quadrupled revenue during that window41531. A Series B landing in the same calendar quarter as a company's public launch is the venture market's bluntest statement of conviction: Bessemer is not underwriting a business, it is underwriting a category.

The round, and the compression it represents

Dealroom's account of the financing notes that Onyx left stealth with the earlier $40 million and that chief executive Maxim Bar Kogan described post-stealth demand as outpacing expectations, adding that early traction "allowed us to raise this round much sooner than expected"15. Calcalist, which broke the valuation, frames the $640 million figure as a several-fold jump from the prior mark, though the company itself did not officially disclose it2031. Seedtable's data fills in the arithmetic: a $527 million pre-money against the $113 million check14.

The speed is the tell. In a normal cycle, a company exits stealth, spends a year or two proving repeatability, then raises its B. Onyx skipped that interval entirely, and its investors are effectively paying an option premium on the assumption that agent adoption inside large companies will be steep enough, and messy enough, that someone has to own the oversight layer. The company's own framing is aggressive on this point: Bar Kogan has said that in 2025 less than one percent of enterprise actions were performed by agents, and that autonomous actors will soon take the overwhelming majority of them4. Onyx's blog pushes the timeline further, arguing that whole companies will run as autonomous operations making decisions about capital allocation, energy grids and supply chains, and that the gap between what advanced AI can do and what humans can govern is widening10.

That is a founder's rhetoric, but it is also the investment thesis in plain language, and the syndicate that bought it is unusually cross-ideological: Bessemer and TCV are classic enterprise-software growth houses, Cyberstarts and Team8-style security specialists carry the cyber credibility, and Conviction is an AI-native fund betting specifically on the application layer of intelligence114. The presence of growth-stage money at the B, rather than a seed fund stretching, signals that investors believe this market sizes like a security platform, not like a point tool.

What Bessemer is actually buying

Onyx sits between a company's AI agents and the systems those agents touch, inspecting each attempted action and blocking the ones that look risky38. SecurityWeek describes a centralized platform that lets organizations identify and regulate advanced AI tools across their networks20. The product set, as PitchBook catalogs it, spans agent discovery and cataloging, policy enforcement, real-time monitoring, incident response and integration with existing security stacks17.

The critical detail in the coverage is who Onyx already works with. The company integrates with Claude Enterprise and ChatGPT Enterprise, counts Fortune 500 companies among its customers, and in June Anthropic announced an integration with Onyx aimed at securing enterprise AI adoption2611. A named customer quote in the announcement from Revolut's head of security engineering describes using the platform to evaluate and stop threats from coding agents across multiple service providers4.

The Anthropic integration is, in my reading, the single most important fact in the story. Model providers have every commercial incentive to bundle governance themselves and every reputational incentive to claim their own agents are safe. Instead, Anthropic is treating an outside control plane as something that accelerates adoption of its enterprise product. That validates the neutral-third-party position rather than undermining it, and it is the strongest structural evidence that Onyx's wedge, the layer that spans vendors rather than serving one, can hold.

There are real caveats. The revenue quadrupling is a company-reported number off an undisclosed base, four months out of stealth, and none of the coverage independently verifies it36. The "AI control" category is also being priced on a story rather than a market: Onyx's own announcement leans on the arrival of new models such as Mythos and Fable, described as so capable they were initially withheld from public release, to argue that agents are becoming both more useful and harder to govern6. That is a narrative, not a financial disclosure, and investors are underwriting it at a valuation that already assumes substantial execution.

Where the coverage agrees, and where it doesn't

The core deal terms are consistent across every outlet: $113 million, Bessemer leading, the seven named participants, $153 million total, and the use of proceeds, which is to train the next generation of Onyx's proprietary models and to scale go-to-market operations in the United States and abroad121115.

The divergences are mostly at the edges, but two are worth flagging. First, the investor list is not stable. Seedtable's database adds Dell Technologies Capital, NVIDIA, Team8, SV Angel and Crescent Cove to the syndicate, names absent from the official announcement14. If accurate, that means the strategic and security-specialist money in this round is deeper than the press release suggests, which would strengthen the read that this is a coordinated land grab rather than one firm's conviction. Second, the company's own footprint is described inconsistently: SecurityWeek calls it Israeli, the New York Business Journal and the SaaS News place it in New York and Tel Aviv, Fundraise Insider notes a Delaware registration, and headcount estimates range from about 70 people to more than 80 across Israel, the US and Canada15111415. None of this changes the deal's meaning, but the discrepancy on investors is the kind of gap that matters if the round turns out to be more oversubscribed than presented.

The bigger funding signal

Zoom out and the round looks less like an outlier and more like the leading edge of a wave. DevX's summary of the announcement places it against a broader run of capital flowing into AI control and governance vendors, as enterprises deploy automated decision tools across finance, healthcare and public services and face pressure to prove those systems act as intended9. Enterprise DNA's analysis is the sharpest version of the buyer's view: budget for agent-security tooling now, because enterprises are already paying for it and will eventually expect vendors to bake it in3.

That last clause contains the existential risk. If agent governance becomes a checkbox inside Claude Enterprise, ChatGPT Enterprise and the incumbent security platforms, the standalone category compresses into a feature and the $640 million price tag ages badly. Onyx's counter is that its value comes precisely from sitting above all of those providers at once, which only works if the enterprise stack stays genuinely multi-vendor.

My own read is that Bessemer made a defensible bet at an aggressive price. The near-term demand is real and measurable in Onyx's reported growth; the medium-term defensibility rests on being the default control point before either the model providers or the CrowdStrike-class incumbents absorb the function. What the $113 million actually buys, then, is time and position, a two-year-old company now capitalized to train its own models and build a global sales motion while the category is still being named115. The wager embedded in the round is that agent actions inside the enterprise go from under one percent to dominant within the funding cycle. If that inversion happens, this round will look cheap. If it stalls, Onyx will have raised a Series B on a story that the market took two extra years to confirm, and its investors will have paid a very full price for the option.