Onyx Security — $113M Series B for an AI-agent control layer (SecurityWeek, July 2026)
Tag: S-2026-07-29-onyx-security-series-b Type: article (independent trade-press coverage of a vendor funding announcement, fetched in full) Author(s): SecurityWeek News Date of source: 2026-07-30 (published); funding announced by Onyx on Wednesday 2026-07-29 Date ingested: 2026-08-07 Authority weight: medium — independent outlet, funding facts consistent and specific, but valuation is second-hand (“reportedly”) and all capability claims are explicitly attributed to the company Raw file: /_raw_sources/S-2026-07-29-onyx-security-series-b.md
What it claims
Onyx Security, an Israeli company founded roughly two years ago, announced a $113 million Series B led by Bessemer Venture Partners, with Cyberstarts, TCV, Conviction, FirstMark, Vintage, QuantumLight and G Squared participating. Total funding reaches $153 million; the round reportedly values the company at an estimated $640 million, which Onyx did not officially disclose. Proceeds are earmarked for training its proprietary models and scaling go-to-market.
The product is described as a centralised platform for identifying and regulating advanced AI tool use across an organisation’s networks, addressing “the risks introduced by highly capable AI agents that are granted access to critical corporate systems without built-in accountability or oversight”. According to the company, it uses proprietary models to track an AI agent’s decision-making process at every step, enabling intervention to “rectify any unintended or malicious behavior as it happens” across SaaS, cloud and endpoint environments. The stated outcomes are finding shadow AI implementations, enforcing real-time safeguards against threats such as prompt injection, and meeting regulatory compliance standards. Hila Zigman, a general partner at investor Cyberstarts, frames the category claim: “Onyx is building the control layer that makes enterprise AI adoption possible at scale… one of the defining security categories of the coming decade.”
Notable quotes
- “The solution is designed to address the risks introduced by highly capable AI agents that are granted access to critical corporate systems without built-in accountability or oversight.” (article body)
- “According to the company, the technology relies on its own proprietary models to track an AI agent’s decision-making process at every step, enabling it to intervene and rectify any unintended or malicious behavior as it happens…” (article body — note the attribution)
- “As AI agents become embedded in critical business workflows, enterprises need a way to ensure they operate safely, predictably, and within policy.” (Hila Zigman, Cyberstarts — investor in the round)
What’s speculative vs. asserted
Asserted: the round size, lead and participating investors, cumulative total, company origin and approximate age, and the intended use of proceeds. Explicitly hedged in the source: the ~$640M valuation (“reportedly”, “estimated”, not disclosed by Onyx, attributed to Calcalist). Company-asserted and unverified: per-step agent decision tracking, real-time intervention and rectification, shadow-AI discovery, prompt-injection safeguards, and the claim that the platform helps “meet regulatory compliance standards” — which names no standard at all and is the sort of unqualified compliance claim this scan treats as marketing rather than fact. The “defining security category of the coming decade” line is an investor in the round and is maximally self-interested. No customer is named in any sector.
Topics this feeds
- AI Governance Platforms — third security-market funding event in roughly ten days directed at the AI-agent inventory-and-enforcement layer, after Neo ($100M, 20 Jul) and Hush Security ($30M, 28 Jul).
Open questions raised
- Onyx is not on Paul’s watchlist and is security-led, not governance-led. The scope tension logged on 2026-07-30 — whether security/identity-capitalised agent-control vendors belong in an AI-governance scan — recurs here and remains unresolved, not settled by this ingestion.
- “Track an AI agent’s decision-making process at every step” is a strong claim about model introspection. What is actually observed — inputs and tool calls, or reasoning — and would either constitute EU AI Act Art. 12 records? [not addressed by the source]
- Roughly $243M has entered this layer from security investors in ten days while the Gartner-defined AI-governance platform category has raised nothing comparable in the same window. Does the AI inventory of record end up owned by security budgets rather than governance functions, and what does that do to second-line independence? [inference]