Kevin Mandia's Armadin Raises $255.5M at $2.5B Valuation

I've read a lot of cybersecurity funding announcements. This one made me put the coffee down. Twenty-six thousand AI hackers — not a metaphor, not a marketing slide, an actual swarm of autonomous agents — attacked one company over three days this summer and found 38 ways in. The company knew they were coming. It had paid to be attacked. And by the attacking startup's own account, it was a well-defended business.

The startup is Armadin. It was in stealth until March; now it has announced a $255.5 million Series B funding round co-led by Andreessen Horowitz and Accel. The money takes its total raised to $445 million and the valuation past $2.5 billion. For a company that barely existed last spring, that is either a staggering vote of confidence or — my honest read — a measure of how frightened the market is.

And here's the kicker: Armadin doesn't sell defense. Not the way anyone else does. It sells the attack.

Armadin $255.5M Series B: a16z and Accel co-lead

The money first, because the numbers are clean even if the idea is wild. Andreessen Horowitz and Accel co-led the Series B — Accel also led the Series A, so the conviction runs deep. Bain Capital Ventures and Redpoint are new names on the cap table, and the old guard piled back in: Google Ventures, Kleiner Perkins, Menlo Ventures, In-Q-Tel, 8VC and Ballistic Ventures, according to the company's announcement. Seven months. That's how long it's been since Armadin emerged from stealth in March with $189.9 million in combined Seed and Series A financing. The Palo Alto company's new war chest goes to the usual suspects — scaling the platform, research, model training, go-to-market. The usual list, attached to a very unusual product.

One name explains why the checks got written this fast. Armadin's chief executive is Kevin Mandia — the man who founded Mandiant in 2004 (yes, the company is basically named after him), sold it to FireEye for $1 billion in 2014, and then watched it end up inside Google in a $5.4 billion deal in 2022. For two decades, Mandia has been the man who gets the 3 a.m. call when a company's worst night begins — the guy in the war room, picking through breached logs while executives wait outside for the damage report. "Kevin has been on the front lines of the most consequential breaches in history, and he has built a team that pairs elite red teamers with world-class AI engineers," said David George, a general partner at Andreessen Horowitz, in the funding announcement. "We invested because we believe Armadin will become the defining security company of the AI era."

Armadin is that instinct industrialized. Instead of one team on a plane — a swarm that never sleeps.

26,000 agents, 17 million actions, 38 ways in

Stop scrolling. This is the number. As the Wall Street Journal reported, Armadin unleashed 26,000 agents in a three-day attack on a customer's business — 26,000 — generating roughly 17 million offensive actions. What came out the other end: 38 validated attack paths. Not vulnerabilities. Paths. Complete routes a real adversary could walk, end to end, into an unnamed company Armadin describes as having solid cybersecurity.

Do the math, because it doesn't feel real until you do. Seventeen million actions over 72 hours. That's about 65 offensive moves a second, every second, for three straight days. No human red team works like that. I've covered pen-test firms for years; the best of them are a dozen people with laptops and a deadline. This is a different animal entirely. That's the point.

I'll admit I got suspicious here. I always do when a company's own funding announcement is also its biggest case study. So I cross-checked. Armadin and TENEX.ai, an agentic security-operations firm, documented the same exercise publicly back in August: a joint engagement against what they called a globally critical institution. Their version of the numbers: 1,300 attack waves, 26,000 agents, approximately 17 million offensive actions, more than 25,000 services probed, 238 security findings — 98 of them significant — chained into those same 38 validated attack paths. The Journal's numbers, Armadin's own announcement, the joint release — all three line up. That doesn't happen by accident.

And the detail that matters more than the fireworks? The agents ran with no privileged credentials, no source-code access, and no whitelisting of the target's security controls, the companies say. They came in from the outside, exactly the way a real adversary would — no cheat codes. (Every action passed through a control layer overseen by a safety model trained on feedback from human security experts, which strikes me as the bare minimum when you're letting 26,000 autonomous hackers loose in a live environment — but at least it's there.) On the defensive side, TENEX.ai's platform chewed through 101,169 alerts and reconstructed the whole campaign across 231 billion raw events — forensic work the companies figure would have taken a five-person analyst team about 2,400 hours. Four months. By hand.

The pitch: scanners are obsolete, offense is the only defense

So what is Armadin actually selling? The phrase they use is "effective autonomous security" — which sounds like consultant-speak until you hear what it means. A swarm of specialized AI agents that reasons like a skilled adversary and roams a customer's attack surface, chaining individually low-severity weaknesses into validated kill chains. The company's announcement describes those chains starting with unauthenticated remote code execution at the perimeter, moving sideways through the network, and ending in full cloud compromise. Security teams see the exact routes an attacker would take, plus the blast radius of each — and, in theory, cut them before anyone else does. Reco raised $55M to secure the AI agents nobody can see.

The company says it's already running these campaigns in production for Fortune 500 enterprises and government customers. Seven months out of stealth. If true — and that's a big if, but it's the claim — the sales cycle here is absurd.

Mandia's argument, as he laid it out to the Journal, is that the old tools haven't just fallen behind — they've failed in a way that's almost embarrassing. "Finding vulnerabilities is no longer difficult, as frontier models do it every day," he said. The result: today's scanners fire off thousands of alerts that swamp cyber teams and bury the serious risks. Armadin's agents chain weaknesses across a company's digital perimeter, into its network and up into the cloud, mimicking how hackers actually move through a target's systems hunting for an opening. Only validated, exploitable risks set off alarms — so defenders stop, in Mandia's phrase, "chasing noise."

The agents are trained by real-world cyber red teams — people who have spent their careers breaking into well-fortified companies and organizations — whose experience feeds the agents' autonomous attack campaigns, Mandia told the paper.

"Offense is uniquely advantaged right now," Mandia said in the company's funding announcement. "AI lets an attacker find and chain weaknesses faster than any human team can respond. The only way to build a defense that keeps pace is to train it against the best offense available, every day. That is what we built."

His investors are happy to say the quiet part out loud. "Against agentic adversaries, the best defense is a great offense powered by AI," said Ping Li, a partner at Accel, which co-led this round after leading the Series A. "That was our conviction when we led Armadin's Series A, and it's even more pressing today." Congress now wants AI labs to disclose Chinese attempts to steal their model weights.

Why VCs paid $2.5B for a seven-month-old startup

A $255.5 million Series B for a seven-month-old company is not normal. The market knows it. Dealroom's data puts this round in the 99th percentile of all-time Series B rounds for U.S. security companies — out of a sample of 1,066 deals. Their reading: AI is compressing startup timelines so fast that seven months of traction now reads like five years of the old world. Nvidia just bankrolled a $668M bet on Taiwan's AI supply chain.

Two forces are doing the compressing. First, frontier AI models have collapsed the gap between a vulnerability being disclosed and a working exploit existing — "as frontier AI models compress the time between vulnerability disclosure and working exploit, periodic penetration tests cannot keep pace," Armadin's announcement puts it. They're right about this one. Second, the threat has gone agentic, and we're not talking hypothetically. Just this week we covered how AI agents tried to hack Canada's national archives — actual SQL injection payloads fired at a government database — and how researchers used Claude to write exploit code that breached OpenAI's own forum. The attackers are already autonomous. Mandia's pitch is simply that the defenders have to be too. Hard to argue with the premise when the evidence keeps arriving weekly.

Armadin $255.5M: the numbers, answered

What did Armadin just raise? $255.5M in Series B funding, announced October 1, valuing the company past $2.5 billion — bringing total funding to $445M, seven months after the $189.9M combined Seed and Series A rounds in March.

Who is Kevin Mandia? Armadin's founder and CEO. He founded Mandiant in 2004, sold it to FireEye for $1 billion in 2014, and FireEye's Mandiant business was then acquired by Google for $5.4 billion in 2022.

What are the 26,000 AI agents? An autonomous red-team swarm trained by real-world red teams, attacking with no privileged credentials, no source-code access and no whitelisting of the target's controls.

What are "38 validated attack paths"? Complete end-to-end routes a real adversary could walk — from unauthenticated remote code execution through lateral movement to full cloud compromise — not just isolated vulnerabilities.

Why does the FTC care? The same week, the FTC opened an industry-wide probe into rogue AI agents, with the chair warning developers who set agents loose on cybersecurity tests could face liability.

Our read: the friendly cyberattack business has arrived — and so has the backlash

Let's be blunt about what Armadin has built: a weapon you point at yourself, on purpose, for money. That is a colossally ambitious bet, and also a completely rational one — because the alternative, the shoddy status quo of scanners firing thousands of alerts at exhausted analysts while frontier models discover new vulnerabilities daily, is already failing. If your security stack can be beaten by an agent that costs pennies per hour to run, the only honest move is to hire the agent first and let it show you how.

The skepticism writes itself. Thirty-eight "validated attack paths" at a well-defended company is the kind of claim that makes a CISO either reach for the checkbook or for the salt shaker, depending on how much of the validation they get to watch. Armadin's answer — no inside access, no whitelisting, and only exploitable risks trigger alarms — is the right answer. Whether it holds up across hundreds of customer environments instead of one showcase exercise is the entire bet behind that $2.5 billion valuation.

And the timing is, to put it mildly, provocative. Armadin is scaling up agent swarms for offense — friendly offense, but offense — in the same week that the FTC opened an industry-wide probe into rogue AI agents, with compelled executive testimony on the table and the FTC chair warning that developers who set agents loose on cybersecurity tests should be liable for the harm they cause. Mandia's response to that world is not to slow down; it's to argue that 26,000 agents attacking your company is the safest thing that could happen to it. Visionary or reckless? The next two years of breach headlines will tell us.

Sources: The spine of this story is same-day reporting from Reuters (Anzar Mehraj) and the Wall Street Journal, with SecurityWeek, Help Net Security and Channel News Asia filling in the details, Dealroom supplying the funding comps, and Armadin's own PR Newswire releases — the Series B announcement and the joint Armadin/TENEX.ai exercise report — providing the company's side.

By Daniel Rao, TechPulse USA — covering the companies, models, and money moves shaping artificial intelligence. We read the fine print so you don't have to.

Members of an OPFOR red team during the Cyber Shield 2016 cyber exercise at Camp Atterbury, Indiana. Armadin says its AI agents are trained by real-world red teams who have spent their careers breaking into well-fortified organizations. Photo: U.S. Army / Sgt. Stephanie Hargett (public domain).

Rows of server racks in a data center — exactly the kind of production infrastructure Armadin's agent swarms attack for customers, mapping how a real adversary would move from perimeter to cloud. Photo: Carl Lender / Wikimedia Commons (CC BY 2.0).