$255.5 million. That’s what a cybersecurity startup called Armadin pulled in a single funding round announced on October 1, 2026, pushing its valuation past $2.5 billion. For context on how fast that happened: the company’s total funding to date sits at $445 million, meaning this one round accounted for more than half of every dollar it has ever raised.
So what does Armadin actually do to earn that number? It runs a swarm of AI agents that behave like hackers, probing systems to find security weaknesses before real attackers do.
If you’ve been reading this site for a while, you know I like to pause on sentences like that one. “A swarm of AI agents that act like hackers” is the kind of phrase that sounds either terrifying or meaningless depending on how familiar you are with the words. Let’s fix that.
What a swarm of agents actually means
An AI agent isn’t a chatbot. The difference matters. A chatbot answers your question and waits. An agent is given a goal and then takes actions on its own to reach it: clicking things, running commands, reading results, deciding what to try next.
A swarm is what you get when you run many of these agents at once, often with different jobs, working the same problem in parallel. Think less “one brilliant assistant” and more “a crowd of tireless interns, each poking at a different door.”
Applied to security testing, that framing clicks into place pretty quickly. Finding weaknesses in software has always been a numbers game. Real attackers don’t succeed because they’re geniuses. They succeed because they’re patient and they try a lot of things. A human security tester, however skilled, has eight hours in a workday and one set of hands.
Agents don’t. That’s the whole pitch.
The part that makes people uncomfortable
I want to name the obvious reaction, because I had it too: isn’t building AI that hacks things a bad idea?
This is where a distinction from the security world helps. Hiring people to attack your own systems is a long-established practice. It has a name, penetration testing, and companies pay good money for it. The logic is simple. You’d rather find out your window is unlocked from someone you hired than from someone who climbed through it.
What’s changing is who does the climbing, and how often. If a human team can test your systems once a quarter, and a swarm of agents can test them continuously, that’s a genuine shift in how often you learn bad news about yourself. For most organizations, learning bad news faster is the entire goal.
The uncomfortable truth underneath is that attackers get the same tools. Agents that are good at finding weaknesses are good at finding weaknesses regardless of who’s holding the leash. A lot of the money flowing into this corner of the industry is, I’d argue, less about optimism and more about not wanting to be the slower side of that race.
Reading the investor list
The round was co-led by Andreessen Horowitz and Accel, with new money from Bain Capital Ventures and Redpoint. Existing backers Google Ventures, Kleiner Perkins, Menlo Ventures, and In-Q-Tel also joined in.
That last name is the one I’d linger on. In-Q-Tel is the investment arm associated with the US intelligence community, and it tends to show up around technology that national security agencies want to understand early. Its presence on a cap table isn’t a guarantee of anything, but it does tell you the problem Armadin works on isn’t considered niche.
The rest of the list reads like a who’s-who of firms that have been writing large checks across AI generally. Which raises a fair question: is this a bet on security, or a bet on agents?
Probably both, and that’s the interesting part. Security testing is a near-perfect showcase for agents because the work is repetitive, measurable, and valuable. You can tell whether an agent found a real weakness. Compare that to the fuzzier promises made about AI assistants in creative work or strategy, where success is hard to define and easy to overstate.
What I’d watch next
Armadin says the money goes toward scaling its platform, strengthening research, and expanding its sales efforts. Standard stuff for a Series B, and not much of a signal on its own.
The more useful signal for the rest of us is the pattern. When agents get deployed in a domain where results are plainly verifiable, they tend to stick. Security is one of those domains. If you want a preview of where agents genuinely earn their keep rather than just generating demos, watch the places where someone can check the homework.
A swarm of software pretending to be burglars, valued at $2.5 billion, is a strange sentence to end on. It’s also a reasonable description of where this technology is finding its footing first.
🕒 Published: