The Biggest Shift in Cybersecurity in 30 Years
Summary
Kevin Mandia returned to company-building because AI is overturning the security landscape after his 30 years in security. He says he felt compelled by Armadin’s founding team and the chance to build what every company needs: “I don’t want to miss out on the changes brought about by AI” when “everything I did before is outdated.”
AI turns cyber offense from a resource-constrained “sniper shot” into a faster, broader “swarm of drones.” Armadin has already coordinated 25,000 agents; Mandia says what AI does in a microsecond could require 70 humans and may be impossible for them. Open models are “already good enough,” with criminal adoption likely to accelerate once attackers can access GPUs anonymously.
Armadin aims to replace periodic penetration testing with continuous, exploit-verified pressure against customer networks. Its “hyperattack” maps assets into a metadata replica, polls that footprint like a heartbeat, and attacks whenever the network or threat environment changes. Since January 2026, it has found more than 90 zero-days across Fortune 500 customer environments, usually reporting remote code execution in a DMZ within 48 hours.
The product thesis extends from finding vulnerabilities to autonomously containing them before attackers arrive. Armadin Blue would turn offensive findings into compensating controls across EDR and firewalls, ideally converting a five-minute intelligence advantage into immediate protection. Mandia’s trade-off is blunt: “I would rather have a failed patch that stops an attacker” than allow an unknown intruder arbitrary network access.
Human-paced detection and response cannot survive agentic attack speed, although layered defenses remain necessary. Mandia expects AI to drive prevention, detection, and response as those phases compress toward one instantaneous cycle; humans will supervise exceptions rather than remain tactically “in the loop.” Near term, AI favors offense, but defensive AI trained against offensive AI should ultimately become more thorough than humans inspecting packets.
Cyber performance may converge faster across open and closed models than in other AI workloads. Across 20 real-world attack chains, no tested model completed more than eight; the most advanced closed models reached exploitable findings faster, but open models reached the same endpoint when allowed to run longer. The discussion points toward attack expertise, evaluation, cost, orchestration, and containment—not merely model access—as durable advantages.
George calls the security shift a “tsunami,” and Mandia says it demands venture-scale speed without abandoning enterprise execution. Armadin has an “80 mph tailwind” but must finance sales, brand, international expansion, and second and third acts; it wants to beat Wiz’s cited path to more than $100 million of ARR within 18 months. With technical advantages potentially copied in six months, the operating moat is still: “Find a customer, make them happy, repeat.”
Deep dive
1. AI pulled a reluctant founder back into the arena
Mandia says returning was not an expression of lifelong entrepreneurial hunger: in his 50s and already in venture, he had not expected to start again. Meeting David Slater, Travis Lam, Evan Peña, and the wider founding team produced the opposite reaction: “I want to be a part of this.”
The catalyst was the prospect that AI would invalidate much of his 30-year security experience. His conclusion was not nostalgia but urgency: the team had the technical talent, while he could contribute relationships, pattern recognition, and a route into companies that already needed the product.
Armadin Red uses advanced models and offensive AI to prove which risks are exploitable. The planned counterpart, Armadin Blue, turns those discoveries into defenses—the combined ambition is a “force field” in which good offensive AI continuously trains good defensive AI.
2. The sniper-shot era is becoming a swarm
Today’s observed AI attacks remain less capable than Armadin’s internal systems. Mandia describes agents moving from point A to B, then oddly attacking A again: coordinated enough to resemble a drone swarm, but still “a little sloppier, a little louder” than a disciplined human operator.
Scale changes the contest even before models become more creative. Cyber operations involve code and structured processes, so Armadin does not require a model fluent in 400 languages; existing models can already explore far more paths, at far greater speed, than resource-constrained human teams.
Nation-state operations traditionally behave like sniper shots: a narrow campaign against perhaps 30 defense contractors or military targets. AI introduces a doctrinal choice between stealth and comprehensiveness—when to preserve the sniper, when to “burn the tokens” on a swarm, and how much human judgment is needed to keep that swarm subtle.
The strategic consequence is democratization: less-skilled attackers may become materially more successful, while defenders lose attribution clarity. A model-driven assault may reveal its automation without revealing whether “this is a country” or merely one person behind it.
3. Continuous exploitation could replace legacy pentesting
Mandia’s sports framing is the 2000 Baltimore Ravens: an elite defense needs an offensive practice squad capable of truly testing it. “You have no defense unless you have a great offense to stand up to,” and that pressure must be continuous rather than periodic.
Armadin’s “hyperattack” sends agent swarms across every service, route, system, and asset, producing terabytes of metadata and an attacker’s-eye replica of the network. It then polls that replica “almost like a heartbeat,” cheaply identifying changes before concentrating attacks on what is new.
Triggers include a changed application, route, service, or machine; new threat intelligence or models; and deliberate self-audits before events such as board meetings. The objective is to determine exploitability, not merely enumerate CVEs, because many bugs do not provide stable access or remote command execution.
Mandia contrasts that proof with scanners such as Tenable, Rapid7, and Qualys, which he characterizes as hygiene tools. Armadin attacks externally as a black box, without source code, verifies remote execution or data extraction, and has driven customers into real incident-response mode—not “a huge list of vulnerabilities that didn’t matter.”
4. Autonomous defense becomes the mandatory second act
Armadin Blue rejects the posture of saying, “Hey, you’re vulnerable. See you later.” If Red discovers an attack five minutes ahead of an adversary, Blue should immediately inform endpoint EDR, firewalls, and platforms including CrowdStrike, Palo Alto Networks, and Fortinet to create compensating controls.
Mandia compares the first generation to battlefield first aid: bandage the wound and stop the bleeding, then let people design the permanent repair. The controls may initially be crude, but autonomous protection will arrive through security platforms “whether you want to or not.”
The SOC’s prevention, detection, and response windows will keep narrowing until the sequence is effectively instantaneous. Attackers inside a network can perform “a thousand things at once”; leaving a person in that tactical cycle is too slow, like watching a balloon burst after the whole defensive apparatus has already failed.
George’s pushback—that layered, “onion” protection still matters—holds. Mandia agrees the force field will eventually miss something, requiring traps that detect illicit access and lock systems immediately, but those backup stages must also operate autonomously.
5. Safe cyber agents require experts who know how attacks fail
Asked about the Hugging Face case, Mandia’s broader diagnosis is that every new technical modality leads defenders to underestimate the adversary—in this instance, the model itself. AI researchers may lack security instincts, while security specialists may not understand AI well enough to “cage this thing.”
Armadin assumes each control can fail: host-level isolation, a hypervisor, proxies, passive inspection of every request, and deterministic prohibitions all overlap. Interestingly, agents are most often stopped not for dangerous behavior but because they are “chasing the wind” and wasting money.
The constraint cannot become absolute. Mandia argues, “Don’t cage it too tightly”: classifiers trained with human thumbs-up and thumbs-down judgments should monitor inputs and outputs, while genuinely ambiguous actions stop and escalate to a person so the system retains useful creativity.
Armadin’s 20-chain evaluation exposed model limits: no open or most advanced closed model completed more than eight chains. Closed models found exploitable vulnerabilities faster, but open models caught up with more runtime; Mandia says humans found most of Armadin’s 90-plus zero-days while AI handled about 90% of routine work, although the latest few were found by the technology.
6. The company-building moat is speed disciplined by customers
Mandia casts Armadin as a third intelligence wave. Antivirus was a “Maginot Line”; Mandiant learned from victims after inevitable breaches; Armadin’s premise is that companies should discover their own exploitable weaknesses before waiting for a defense contractor or another victim to be compromised.
Unlike self-funded, profitable Mandiant in 2004, Armadin needs capital to exploit its “80 mph tailwind,” build distribution, and prepare successive acts. Mandia identifies four changed requirements: financing, speed, brand value, and go-to-market development.
Enterprise sales still depend on people, especially as the technology changes rapidly. Sales teams therefore need weekly technical training, leaders who can scale institutional processes, and a feedback loop that moves customer lessons directly to engineers; the CEO’s job is to “completely hide the chaos” from employees.
Competition is now unavoidable, and Mandia says what a company does today could be replicated in six months. His durable formula is deliberately unglamorous—win demanding banks, retailers, airlines, and other credibility-bearing customers, then “find a customer, make them happy, repeat.” George adds that every technology stack in cybersecurity could change within two years.