Onyx Security

• Platform vendors such as Palo Alto Networks, CrowdStrike, and Microsoft could fold AI agent discovery and runtime governance into suites enterprises already buy, undercutting a standalone Onyx purchase before it confirms a reference customer outside its own page.
• A competitor, Bold Security, surfaced the same week with the same $40 million, and several funded startups address the same agent security slot, so Onyx competes for the same buyers without a public differentiator yet.
• No paying customer organization or case study with deployment scope is visible in the reviewed sources, so the named CISO testimonials on the homepage reduce but do not close the reference-proof gap a procurement team raises.
• Identity providers such as Okta and Microsoft Entra could extend agent identity and access governance into the control-plane role Onyx is selling, removing the third-party budget line.
• The differentiation depends on a supervisory Guardian Agent that no third party has evaluated, so a published benchmark or a named-customer outcome is what would convert the team and funding signals into proof.

Onyx Security treats the autonomous AI agents an enterprise runs as systems it must discover, watch, and constrain before they act on critical infrastructure. CEO Maxim Bar Kogan frames the problem as the loss of control when something the company depends on can make decisions it cannot fully understand. The buyers, per the homepage, are the security, governance, and AI teams adopting agents across departments.

Independent reporting corroborates the gap beyond vendor marketing. SiliconANGLE described organizations deploying AI across their infrastructure and creating new attack surfaces and operational risks that traditional security tools struggle to address. That framing places the pain Onyx sells against in the words of a non-vendor source rather than the company's own pitch.

The risk Onyx emphasizes is autonomous action on real systems. An agent given access to critical systems can make a serious mistake or be compromised through a crafted prompt, which turns a reasoning error into a live operational incident. Discovery and runtime control are meant to catch that before it reaches a downstream system. [s2, s5, s4]

The Onyx platform centers on a supervisory agent rather than a single control point. SecurityWeek describes the Onyx Guardian Agent continuously discovering AI agents across cloud, endpoints, code, and SaaS deployments, then monitoring them and approving or correcting their actions to enforce security and governance policies. The catalog entry adds that it covers sanctioned and shadow AI and monitors prompts and agent actions in real time.

The runtime layer is where the product claims to act. The Guardian Agent can block an unsafe action, require human approval, narrow an agent's scope, or redirect it toward a safer path before anything reaches a downstream system, per the launch announcement. The same announcement positions the Guardian Agent as the supervisory layer overseeing every other agent in the environment.

Public technical depth stops at that description. So soon after launch there are no architecture docs, no open-source components, and no third-party evaluation in the pages reviewed, so a buyer evaluating the Guardian Agent would start from sales conversations rather than testable artifacts. [s4, s2, s6]

Onyx competes in AI agent security against both independents and the platforms building toward the same controls. SiliconANGLE reported that Bold Security surfaced the same week with the same $40 million. Several other funded independents cover overlapping discovery, monitoring, and runtime enforcement for the enterprise AI buyer.

Onyx's visible differentiator is its founding team and its supervisory Guardian Agent. Bar Kogan came from Unit 8200 and Gil Elbaz reported to the CTO of Nvidia, per CTech, and the company positions a Guardian Agent that oversees other agents that rivals would have to reproduce. That pedigree is the asset a bundled competitor cannot quickly copy.

The structural risk is who owns the buyer relationship. Platform vendors can bundle agent governance into deals an enterprise already signs, and identity providers can extend agent access governance from the credentials they already manage. Onyx's independence is both its neutrality pitch and its exposure. [s5, s3, s4]

Onyx points to enterprise traction that is now partly named. The launch announcement says Fortune 500 companies already use the control plane to govern their agents, and SecurityWeek reports it is working with multiple large companies, which is more than design-partner language. The homepage closes part of that gap with named CISO testimonials shown beside customer logos, including Eli Edelkind beside the CAVA logo.

Investor conviction is the clearer signal. Conviction, founded by Sarah Guo, led the $35 million Series A and Cyberstarts backed the earlier seed, per CTech, which is capital betting on the motion rather than buyer proof. The launch drew coverage in SiliconANGLE, SecurityWeek, and CTech, which builds awareness at the moment of entry.

The motion is early and direct. Onyx routes prospects to a demo request and plans to spend the raise on product, engineering, new AI models, and go-to-market, per CTech. The remaining proof gap is a named customer organization or case study with deployment scope, not merely named individual testimonials. [s2, s3, s10]

The founders pair offensive-security and AI-research backgrounds. CTech reports that CEO Maxim Bar Kogan is a security leader and technology executive who served in Unit 8200 of the IDF, the country's signals intelligence unit. Co-founder and Chief AI Officer Gil Elbaz is a veteran AI researcher who previously reported to the CTO of Nvidia and served in an IDF AI research unit.

The pedigree is in-domain and verifiable through press. What is not yet public is a prior exit or a sustained research publication record under the Onyx name, so the credibility comes from background rather than a track record the company has built together.

The team has scaled quickly. Onyx reached about 70 employees across Israel, the US, and Canada in roughly 18 months of building, per SecurityWeek and the launch announcement, which shows the founders can recruit at the pace the funding implies. [s3, s4, s2]

Onyx sells a product that inspects enterprise AI traffic, so the handling of that data is the trust question a buyer raises first. The launch materials describe runtime monitoring of prompts and agent actions and enforcement of governance policies, which means proprietary AI activity flows through the control plane.

Onyx publishes a SafeBase trust center at trust.onyx.security that lists SOC 2 Type 2, ISO/IEC 27001:2022, GDPR, and HIPAA compliance, with SOC 2, ISO 27001, HIPAA, and penetration-test reports available for download after a security review request. The compliance frameworks are stated on the page, while the underlying reports are access-gated rather than open, so a procurement team can confirm the attestations exist but still works through Onyx to read them. The page does not state the SOC 2 reporting period or the trust-service criteria its report covers, so those scope details remain a sales-cycle question. A responsible-disclosure policy is also listed, answering the vulnerability-reporting question a security review raises. [s9, s4, s2]

Add analyzed competitors to compare them side by side with Onyx Security.

Onyx aims the control plane at the enterprise teams standing up AI agents across departments. The homepage names the buyer as security, governance, and AI teams that need to protect, optimize, and measure AI agent adoption across the enterprise. That is a defined buyer with a budget owner rather than a generic enterprise audience, and the launch framing ties the buyer to a dated trigger, the spread of autonomous agents into production in 2026.

The segment skews to the large, regulated enterprise. CEO Maxim Bar Kogan says the control plane is already used by Fortune 500 companies to govern their agents, and SecurityWeek reports Onyx is working with multiple large companies, which points the motion at negotiated enterprise deals rather than self-serve buyers. A Cyberstarts partner frames the same target, calling the team well-positioned to win AI security for large enterprises.

The segmentation risk is breadth inside that enterprise. Onyx pitches observability, security, governance, and orchestration to security, governance, and AI teams at once, three buyers with different priorities, so the open question is which of those teams opens its budget. A company this soon after its March 2026 stealth exit still has to prove which department owns the budget line it is selling into.

The Onyx platform centers on a supervisory agent rather than a single inspection point. The launch announcement describes a single, unified enforcement layer that continuously discovers every AI asset across the enterprise, agents, models, applications that embed AI, and MCP-connected tool ecosystems, and turns governance policies into enforceable controls that operate at runtime. SecurityWeek adds that the system discovers AI agents across cloud, endpoints, code, and SaaS deployments and then monitors and corrects their actions.

The Guardian Agent is the capability Onyx leans on hardest. The company positions it at the heart of the platform, overseeing all other AI agents and able to block an unsafe action, require human approval, narrow an agent's scope, or redirect it toward a safer path before anything reaches a downstream system. SecurityWeek reports that Onyx uses supervisory agents and proprietary AI models to understand AI reasoning and react in real time, which is the differentiator a discover-and-monitor rival would have to reproduce.

Public technical depth stops at that description. So soon after the March 2026 stealth exit there are no architecture docs, no open-source components, and no third-party evaluation of the Guardian Agent or the proprietary models in the pages reviewed, so a buyer assessing how the supervision actually works would start from sales conversations rather than testable artifacts. The proprietary-models claim raises the replication cost, but no benchmark lets a buyer confirm it.

Onyx runs a direct enterprise motion anchored by founder visibility and investor credibility. The CEO wrote the launch announcement and routes prospects to a single call to action, a demo request, which fits a company still proving a repeatable sales motion rather than one running a built-out quota-carrying team. Founder-fronted selling at this stage is the expected signal, not a weakness.

The traction Onyx shows is now partly named. The launch says Fortune 500 companies already use the control plane and SecurityWeek reports work with multiple large companies, stronger than design-partner language, though those accounts stay unnamed in independent reporting. The homepage closes part of that gap with two named CISO testimonials, one from Eli Edelkind shown beside the CAVA logo and a second from Kyle Weckman, whose employer the page does not name, named-CISO testimonials rather than anonymous quotes. They are still vendor-displayed, so the proof a rival cannot also claim is the same buyer speaking on the record outside Onyx's own page.

Investor conviction carries the rest of the early signal. Conviction, the firm Sarah Guo founded, led the 35 million dollar Series A and Cyberstarts backed the seed, and a Cyberstarts partner went on the record calling the team well-positioned to lead AI security for large enterprises. That is capital betting on the motion, and a named enterprise customer speaking publicly is the signal that would convert launch attention into proof of deployment.

Onyx publishes no prices, which itself signals the deal it wants. The homepage and launch materials route every prospect to a demo request rather than a pricing page, the negotiated-enterprise default for a vendor selling into Fortune 500 buyers. Hidden pricing fits the large, custom deal the segmentation targets, and it is the norm for the AI agent security cohort at this stage.

The product framing hints at what Onyx believes the buyer pays for. The homepage describes opt-in governance coverage and policy controls defined in natural language, and the company sells the ability to protect, optimize, and measure AI agent adoption, which points the value toward the scale of an enterprise's agent fleet rather than a seat count. Charging by the governed agent or the enterprise would tie the price to the exposure a security team is trying to contain.

What a buyer cannot do yet is compare. With no published price and no disclosed packaging in the pages reviewed, an enterprise evaluating Onyx against the several funded rivals in the same slot has no public anchor, so the pricing conversation happens entirely inside the sales cycle. For a first priced round, that opacity is conventional but leaves the unit of value unstated in public.

Onyx delivers control as a runtime layer woven through the enterprise's existing tooling rather than a standalone scanner. The AI Defense Matrix Catalog's Onyx entry describes integration with browser, AI platforms, CNAPP, SASE, and EDR discovery sources so nothing is invisible, which means Onyx rides the discovery feeds an enterprise already runs to find agents across its environment. The enforcement runs in real time, protecting AI systems and the data they touch without disrupting workflows, per the same entry.

The runtime posture is the operational heart of the product. The Guardian Agent acts before an action reaches a downstream system, blocking, pausing for human approval, or redirecting it, which places Onyx in the live path of agent decisions rather than in a periodic audit. That continuous interception is what turns governance policy into enforcement, and it is also what raises the bar on the platform's own reliability and latency.

The operational depth a buyer can verify stays shallow in public. The pages reviewed describe discovery, monitoring, and enforcement but expose no public documentation, status page, or service-level commitment, so a security review would have to probe latency, failure modes, and how Onyx handles the agent traffic it inspects through direct conversation. For a product sitting in the live path of enterprise AI, those operational answers are the ones procurement will press hardest.

Onyx sells a product whose job is to make an enterprise's AI auditable, so its own data handling is the trust question a buyer raises at the outset. The platform inspects every prompt, response, and agent action to enforce policy, which means proprietary AI activity flows through the control plane. The homepage positions governance, enforcement, and auditability as the value, and a displayed CISO testimonial frames Onyx as what lets a security team manage AI risk across copilots and other tools.

Regulatory alignment is the company's trust narrative. Onyx says it helps buyers satisfy AI security standards and regulatory requirements with opt-in coverage and policy controls defined in natural language. That positions the product as the layer that produces the compliance evidence legal and risk teams require, which is a credible pitch to a regulated enterprise.

The attestations a procurement team wants are published. Onyx runs a SafeBase trust center at trust.onyx.security that lists SOC 2 Type 2, ISO/IEC 27001:2022, GDPR, and HIPAA compliance and offers SOC 2, ISO 27001, HIPAA, and penetration-test reports for download after a security review request. The frameworks are stated openly while the reports sit behind an access gate, so a buyer can confirm the attestations exist but reads the evidence through Onyx, and the page lists a responsible-disclosure policy for vulnerability reporting. What the page does not state is the SOC 2 reporting period or the trust-service criteria the report covers, so the depth and scope of the audit remain a sales-cycle question rather than a settled fact for a company this soon after its March 2026 stealth exit.

Onyx is building a control plane that sits above the AI an enterprise already runs rather than a feature inside one vendor's stack. The launch announcement positions a single, unified enforcement layer across agents, models, applications, and MCP-connected tools, and the catalog describes integration with browser, AI platforms, CNAPP, SASE, and EDR sources. That neutral, cross-vendor position is the pitch a platform vendor's bundled control cannot fully match.

The ecosystem exposure runs in two directions. Onyx depends on the discovery feeds of the security platforms it integrates with, which is leverage those platforms hold, and the same control-plane role is one the platform vendors and identity providers could ship themselves into deals an enterprise already signs. Gili Raanan calls AI agents the biggest enterprise platform shift since cloud on the Onyx homepage, which is the prize that also draws the incumbents toward the same control point.

The structural advantage Onyx claims is the supervisory layer itself. A Guardian Agent that oversees every other agent using proprietary models is a harder position to bundle than a discovery scan, and no marketplace listing or named partner program a copycat could match by writing software appears in the public record yet. The bet is that owning the runtime control point compounds faster than a platform vendor can extend its suite into it.

Onyx pairs an offensive-security founder with an AI-research founder. CTech reports that CEO Maxim Bar Kogan served in Unit 8200 of the IDF, and co-founder Gil Elbaz is a veteran AI researcher who previously reported to the CTO of Nvidia and served in an IDF AI research unit. That split maps directly onto the product, security supervision built on proprietary AI models.

The pedigree is in-domain and verifiable through press, which matches the same-asset peers in this category. What is not yet public is a prior startup exit or a sustained research publication record under the Onyx name, so the credibility comes from background rather than a track record the two founders have built together as operators.

The team has scaled at the pace the funding implies. Onyx reached about 70 employees across Israel, the US, and Canada in roughly 18 months of quiet building, per SecurityWeek and the launch announcement, which shows the founders can recruit quickly. Converting that headcount into named, referenceable enterprise deployments is the execution test the next year sets.