Cybersecurity

The Pen Test Moved Into Production and Stayed

Horizon3 raised 250 million dollars at a valuation above 2 billion for NodeZero, which runs penetration tests continuously against live production systems. The company opened an Amsterdam office in June.

CybersecurityBy Servola Tech Desk2026-08-034 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Key takeaways

Horizon3 raised 250 million dollars at a valuation above 2 billion, up from roughly 650 million at its Series D in May 2025. The product it funds runs penetration tests continuously against live production systems rather than once a year against a sample.
The company's proof point is 310,000 production security tests with zero disruptions. That is the vendor's own count measured against the vendor's own definition of a disruption, which is worth establishing before it appears in a board paper as an external fact.
Chief revenue officer Matt Hartley puts the traditional comparison plainly: a conventional engagement examines 2 to 3 percent of a network once, while continuous testing covers all of it. On coverage that argument is strong and an annual pentest genuinely does sample.
The change is not to your security budget, it is to your change-control record. An autonomous agent acting inside production needs a named human authoriser you can produce on request, and most approval workflows were written for scoped engagements with signed rules of engagement.
Horizon3 established an Amsterdam office in June 2026 and is funding a deeper EMEA push, so European buyers will see this pitch during the current budget cycle. Settle the authorisation question before the procurement conversation, not during it.

What 250 million dollars just bought

Horizon3 closed a 250 million dollar Series E at a valuation above 2 billion dollars, having raised 100 million at roughly 650 million in May 2025. NightDragon and NEA, both existing backers, co-led. The new names span an unusual range for a security round: Acrew, Blue Cloud Ventures, Demeter Group, PSG Equity and Sapphire alongside Singapore's state-linked EDBI and SAIC Ventures, the venture arm of the American defence contractor, with Craft Ventures, Qualcomm Ventures, Ridge Ventures and SignalFire returning. The company is approaching 100 million dollars in annual recurring revenue with 120 percent year-on-year growth and counts roughly 7,200 customers, from managed service providers to Fortune 10 enterprises.

The product is NodeZero, which the company markets as autonomous penetration testing and describes in-house as AI hackers. Founders Snehal Antani and Anthony Pillitiere came out of Joint Special Operations Command, and the pitch carries that lineage: rather than hiring a consultancy to attack a defined scope for two weeks, you run an attacker continuously against everything. Horizon3 says it has spent around 100 million dollars on research and development building automated systems designed to stay predictable and controllable, which is a revealing thing to have to say about your own product.

The coverage argument is the strong one

Matt Hartley, the company's chief revenue officer, makes the comparison in terms that are hard to argue with. A conventional penetration test examines 2 to 3 percent of a network, once. NodeZero scans the whole infrastructure continuously. Anyone who has commissioned an annual test knows the shape of the problem: you pay for a scoped engagement, the scope is negotiated down to what fits the budget and the window, and the report describes the state of a subset of your estate on a fortnight that ended before the findings were written up.

Continuous testing against the real environment fixes a genuine defect, and it is the reason this category is growing rather than a marketing conceit. Hartley's second claim is the load-bearing one: the platform can test live systems without shutting down operations. That is the capability the whole proposition rests on, because a tool that occasionally takes production down is not a tool anybody runs continuously, however good its coverage.

310,000 tests, and who counted them

The evidence offered for that safety claim is 310,000 production security tests with zero disruptions. It is a substantial number and there is no reason to think it false. It is also, unavoidably, the vendor counting its own outcomes against its own definition of what counts as a disruption, and those two facts sit together without contradiction. A test that briefly degraded a service, or that triggered an alert someone spent an hour clearing, may or may not be inside that denominator depending on a definition nobody outside the company has seen.

The practical instruction is small and worth following. When that figure reaches a board paper or a risk committee, it should arrive labelled as a supplier statistic rather than an independent finding, and the procurement conversation should ask what the company classes as a disruption and whether customers can see their own incident counts. A vendor confident in the number will answer both without difficulty. The answer also gives you something to hold the contract against later, which a headline figure never does.

The question is authorisation, not efficacy

The decision this round should prompt is not whether autonomous testing works. It is who authorised the action. A traditional engagement produces a signed rules-of-engagement document naming a scope, a window and a human being who approved both. A continuously running agent produces findings, and the approval that covers it was given once, at procurement, for a class of activity rather than an act. Most organisations' change-control processes were written for the first model and have not been revisited for the second.

That gap becomes concrete under European operational-resilience expectations, where an institution has to be able to reconstruct who did what in a production system and under whose authority. The workable answer is not to avoid the tooling, which would be a poor trade for the coverage it buys. It is to write the authorisation down before deployment: which systems the agent may touch, what it may not do without a fresh human decision, who holds that decision, and how an action taken at three in the morning is reconstructed six months later. Horizon3 opened its Amsterdam office in June and is spending this round on EMEA, so European buyers will face that question inside the current budget cycle.

Frequently asked questions

Does this replace our annual penetration test?

For coverage, largely yes, and that is the honest case for it. For assurance purposes it may not, because some regimes and many insurers still expect an independent third-party assessment on a defined cadence. Check what your specific obligations name before you cancel anything: continuous internal testing and an external attestation are answering different questions.

Is running attack tooling in production actually safe?

The vendor's record says yes at 310,000 tests. Treat that as a strong prior rather than a guarantee, and reduce your own exposure the ordinary way: stage the rollout, start with systems whose failure is survivable, and confirm your monitoring can distinguish the tool's activity from a real intrusion before you widen the scope.

Why does a security funding round matter to a European buyer now?

Because the money is explicitly funding an EMEA push and the company established an Amsterdam presence in June 2026. A well-capitalised vendor tripling its valuation will be in front of European procurement teams this year, which makes now the time to decide your position on autonomous action in production rather than deciding it under sales pressure.

Continuous testing against live systems is a better answer to a real problem than the annual sample it replaces, and the market is pricing it accordingly. What an operator adds to that judgement is a record. The organisations that adopt this well will be the ones that can still name, a year from now, the person who authorised an autonomous agent to act inside production and the boundary they set on it. That document costs an afternoon to write before deployment and cannot be reconstructed afterwards.

Horizon3NodeZeroautonomous penetration testingNightDragonNEAoperational resiliencechange controlEMEA expansionAmsterdamsecurity procurement

A Police Dashboard on 700 Foreigners Sat Open

A police surveillance dashboard in Zhangjiakou held profiles on more than 700 foreign residents and 300 foreign journalists, and it was left unsecured. The vendor patent behind it is the part that should change how you brief staff travelling to China.

4 min read

Cybersecurity

Nobody Can Verify The Chip In Your Security Key

On 6 August the DEF CON badge ships with Baochip-1x, a 22nm part built so the silicon can be checked with infrared light. Every root of trust you already own cannot be checked at all.

6 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles