Imagine this: An AI agent in your engineering team hallucinates a critical code change, pushes it to production, and accidentally exposes customer data to the world. Or worse, a crafty attacker slips in a malicious prompt, hijacking your customer service bot to siphon sensitive info. Sound like sci-fi? It's the harsh reality enterprises are barreling toward in the "agentic era," where autonomous AI agents are infiltrating every corner of business—from code deployment to IT ticketing and customer ops.[1][2]
Enter Onyx Security, the Israeli startup that just burst out of stealth mode on March 12, 2026, armed with $40 million in funding to build the world's first secure AI control plane. Backed by heavy-hitters Conviction (led by Sarah Guo) and Cyberstarts, Onyx isn't just another AI security band-aid. It's a full-spectrum platform designed to discover, monitor, govern, and protect these "digital employees" before they wreak havoc—or get compromised themselves.[3][4]
In this deep dive, we'll unpack why Onyx's launch is a game-changer for the Onyx Security AI agents funding story, how their tech stacks up, and what it means for enterprises racing to scale AI without imploding their security posture. If you're knee-deep in AI tools or plotting your agent strategy, buckle up—this is your roadmap to safe agentic AI adoption. (Pro tip: Pair this with our guide on top AI agent frameworks for the full picture.)
The Agentic Explosion: Why Enterprises Need a Control Plane Now
AI agents aren't tomorrow's tech—they're yesterday's pilot projects going rogue today. These autonomous systems, powered by large language models (LLMs), don't just chat like old-school copilots. They reason, plan, act, and iterate across tools, clouds, endpoints, code repos, and SaaS apps. Think Devin for coding, or custom agents triaging support tickets or automating procurement.[1]
But here's the rub: Agents are non-deterministic. They pull from dynamic contexts, untrusted inputs, and external data, leading to hallucinations, reasoning errors, or prompt injections that traditional firewalls miss. Market data paints a grim picture:
- 80% of enterprises expose sensitive data through unchecked agents.[3]
- 93% run agents with excessive permissions.[3]
- 70% face remote code execution risks via compromised agents.[3]
Broader stats underscore the boom: The agentic AI market hits $9.14 billion in 2026, rocketing to $139 billion by 2034 at a 40.5% CAGR.[5] Yet, 81% of teams have deployed agents past planning, but only 14.4% have full security approval, with 88% reporting incidents.[6]
"Every enterprise is becoming an agent operator—whether they planned to or not," warns Onyx CEO Maxim Bar Kogan. "Agents are given access to the most critical systems... but what are our guarantees they will not make serious mistakes or get compromised?"[1]
Legacy tools? Useless. They're built for static apps and human workflows, blind to agent reasoning chains. Onyx steps in as the "air traffic control" for this chaos, letting security, governance, infra, and exec teams collaborate on one platform.
Onyx Security: From Stealth to $40M War Chest
Founded in 2024, Onyx emerged from 18 months of stealth with a blockbuster $40M total ($5M seed from Cyberstarts in 2024 + $35M Series A led by Conviction).[4] The cash fuels product/engineering expansion, new proprietary AI models, and GTM scaling—with a sales push into the US (CEO Bar Kogan is relocating soon).[4]
Team powerhouse: 70+ employees across Israel, US, and Canada. Co-founders Maxim Bar Kogan (ex-Unit 8200, Mixtiles VP Product/Eng) and Gil Elbaz (ex-Nvidia AI researcher under CTO, IDF AI unit) blend cyber offense and agentic AI smarts. Execs include Rob Witmer (VP Sales), Moshe Kanooni (VP R&D), and board stars like Sarah Guo (Conviction) and Hila Zigman (Cyberstarts).[7][2]
Already live with multiple Fortune 500 customers, Onyx proves traction amid the "AI security race."[3] Their platform? An agent-native control plane securing 137,000+ agents, covering 593,000+ employees, and scanning 10M+ sessions for threats.[3]
Investors rave: Sarah Guo calls Onyx the "control plane all enterprises will need as they scale to thousands of agents." Hila Zigman highlights it as "designed for the AI era, rather than adapting legacy tools."[4]
Inside the Onyx Platform: Guardian Agents and Real-Time Magic
Onyx's secret sauce is the Onyx Guardian Agent—a supervisory AI overlord powered by proprietary models that grok agent reasoning. It doesn't just watch; it intervenes in real-time across SaaS, cloud, endpoints, and code.[2]
Core Features:
- Discovery: Auto-finds sanctioned/shadow AI—agents, models, apps, toolchains (e.g., MCP ecosystems).[3]
- Monitoring: Tracks every prompt, response, reasoning step for anomalies.[1]
- Intervention: Guardian blocks risks, demands human approval, scopes down permissions, corrects paths, or redirects—pre-downstream impact.[2]
- Security & Posture: Hardens against prompt attacks, hallucinations; optimizes compliance (e.g., natural language policies).[3]
- Governance & ROI: Enforces regs, tracks adoption metrics, costs, latency��proving AI value to CEOs.[4]
- Orchestration: Frictionless multi-cloud agent setup.
| Feature | Benefit | Real-World Win |
|---|---|---|
| AI Observability | Visibility into shadow AI | Spots 137K+ agents across 593K users[3] |
| Runtime Protection | Block/correct hallucinations | Prevents data leaks in customer service bots |
| Compliance Engine | Natural language policies | Meets GDPR/SOC2 without custom code |
| Optimization | Cost/latency tuning | Cuts LLM bills 20-30% for infra teams |
Bar Kogan emphasizes: "We define what AI is allowed to do... In code and customer interactions, we already see measurable returns."[4]
Customers like it: "Onyx gives end-to-end oversight... enabling us to double down on AI," says one CISO.[3]
Onyx vs. the Field: Agent-Native Beats Legacy Every Time
No exact twin exists yet, but Onyx carves space from general AI guards (e.g., Lakera for prompt inj, or DLP-focused like Nightfall). Legacy SIEM/SASE? Laughable—they can't parse reasoning chains.
| Aspect | Onyx Security | Traditional Tools |
|---|---|---|
| Target | Dynamic agents (reasoning/actions)[1] | Static apps/human workflows[1] |
| Visibility | Full reasoning traces | Blind to AI internals |
| Intervention | Proactive Guardian AI (block/redirect)[2] | Reactive alerts only |
| Teams | Sec/Gov/Infra/Execs | Sec-only |
| Scale | Thousands of agents | N/A for autonomy |
Onyx wins on unification—think Palo Alto's agent mesh but AI-first. Pairs well with tools like LangChain or CrewAI; check our review of agent builders.
Pros:
- Early F500 traction, scales to 1000s agents.[3]
- Proprietary models decode AI "black boxes."
- Multi-dept ROI: Compliance + optimization.
Cons (fair critique):
- New deps on Onyx's Guardian—meta-risk if compromised (mitigated by its design).
- Enterprise sales cycle; early-stage pricing opaque.
- Agent ecosystem immaturity means evolving standards.
Voices from the Trenches: Quotes That Hit Home
- Maxim Bar Kogan (CEO): “I want to make sure that AI agents are not too independent... This helps the organization understand the return on its AI investment.”[4]
- Sarah Guo (Conviction): “Maxim and Gil saw the need for an AI control plane before the market did... Onyx is the control plane all enterprises will need.”[1]
- Hila Zigman (Cyberstarts): “Onyx [is] building a solution designed for the AI era.”[4]
- CISO Testimonial: “Onyx provides the visibility, governance... to control AI usage, protect what matters, and accelerate the business.”[3]
The Road Ahead: Scaling Safely in 2026
With agent markets exploding ($10.91B in 2026 per Grand View),[8] Onyx positions as the must-have infra layer—like Kubernetes for cloud, but for agents. Expect integrations with Azure AI, AWS Bedrock, or Anthropic soon. For readers: If deploying agents via AutoGen or LlamaIndex, Onyx is your security net—mention it in pilots for vendor perks.
See our deep dive on AI governance tools to stack with Onyx.
FAQ
What exactly is an AI control plane, and why does Onyx call theirs "secure"?
It's a centralized layer for discovering, governing, and securing agents—like a Kubernetes orchestrator but AI-native. Onyx's is "secure" via Guardian agents that intervene on reasoning flaws/attacks, unlike passive monitors.[2]
How does Onyx's funding break down, and who's behind it?
$5M seed (2024, Cyberstarts) + $35M Series A (Conviction lead) = $40M total. Angels from cyber/AI join; funds R&D, team growth (70+ now), US expansion.[4]
Can Onyx handle thousands of agents across Fortune 500 ops?
Yes—proven with 137K+ agents secured, 10M+ sessions analyzed. Features ROI dashboards for exec buy-in.[3]
What's the biggest risk if I ignore tools like Onyx?
Ungoverned agents = data breaches, compliance fails, $millions in fines. 88% already hit incidents; Onyx prevents that at scale.[6]
Ready to agent-up securely? How many AI agents are running wild in your org right now—and what's your plan to wrangle them? Drop your thoughts below!