See pricing
Building Defense for the Agentic Era: Kevin Mandia
The a16z ShowSecurity

Building Defense for the Agentic Era: Kevin Mandia

The founder of Mandiant explains why AI attackers move too fast for humans, how his new company hit Fortune 500 networks with 90+ zero-days this year, and what a startup needs to win a crowded market at AI speed.

October 6, 2026 · 49 min listen · 12 min read · Kevin Mandia
0:00
–:––

Context

Kevin Mandia spent 30 years in cybersecurity and founded Mandiant, which ended up inside Google. He is now founder and CEO of Armadin, a startup that uses AI to attack customers' own networks so they can fix weaknesses first. David George of a16z interviews him about what changes when attacks move from human speed to machine speed. It matters to PMs because it is a live case of a new product category forming under extreme time pressure, with lessons on building a product, a go-to-market motion and a team when the technology shifts every two weeks.

The Big Idea

When attackers use AI, they gain scale and speed that humans cannot match, so the only workable defense is autonomous defense trained by an AI attacker built by the good guys.

Kevin's argument is that no human can stay in the detect-and-respond loop once an AI agent can do in a microsecond what would take 70 people. The same speed shift also changes how a startup must be built and sold.

Key Insights

AI changes the economics of attack

  • Scale: a human attacker used to find one path into a network and had to be selective. AI can probe many paths at once.
  • Speed: Kevin says what AI does in a microsecond would take 70 humans.
  • Why it works for offense: finding vulnerabilities is code, which is a structured language and process. Kevin says that structure is why AI is already great at it, and the model does not need to be the fanciest one.

Weaker attackers become more dangerous

Kevin expects less technical, less successful attackers to look far more successful because of volume. AI is "the equalizer". He also expects attribution to get harder. Defenders will see attacks by models and may not know if a nation or a human is behind them.

Nation-state tactics may shift to swarms

Kevin says nations have historically hacked with a sniper round: deep, targeted attacks on a small list such as 30 defense contractors. AI pushes this toward a drone swarm, which is sloppier and louder but more comprehensive. He says nations are still working out their doctrine, including when to swarm and when to stay quiet. Today AI is not surreptitious unless it has had a lot of post-training and human guidance.

Why open models matter here

Kevin disagrees with the idea of slowing down models to limit cyber risk. His view is that open models are already good enough. What holds back criminal attacks today is cost and the lack of anonymous GPU access. He expects far more criminal AI attacks once anonymous GPUs are easy to get.

Open and closed models tied

Armadin built 24 kill chains that humans had carried out in the real world. No model completed more than eight of them. Kevin says both open-weight and top closed models reached the same eight.

  • Closed models found exploitable risk faster.
  • Open models, left to run longer, reached the same place.
  • The gap that remained was cost, not effectiveness.

Humans in the loop are slow

Kevin's position is that humans cannot stay in the detect-and-respond loop. Prevent, detect and respond will all be run by AI. When Armadin's own agents break in, they spread across a network at a speed he calls shocking, doing a thousand things at once instead of one command at a time. He expects some SOC processes to disappear and every vendor to be rethinking its role.

Humans still find most zero-days

Of Armadin's 90+ zero-days, Kevin says most were found by humans, but those humans used AI for 90%+ of the tedious work. Then a turn came: the last few zero-days were found by the technology itself.

Mental Models & Frameworks

Continuous offense through a heartbeat

Armadin runs a "hyper attack": a swarm of agents maps every service, route and asset on a network, producing terabytes of metadata. Think of it as a metadata twin of the customer's environment.

  • Poll cheaply: check that metadata like a heartbeat for changes, such as a new route, an updated service or a new machine.
  • Attack the change: test only what changed instead of attacking everything all the time.
  • Trigger a full test when: new models or intelligence appear, the network changes, or the customer just wants an audit (Kevin's example: a board meeting tomorrow).
  • Why not constant attack: Kevin says continuous full attacks would be cost-prohibitive and unnecessary.

Red first, then blue

Kevin describes Armadin's plan in acts. Act one is Armadin Red, the best at finding exploitable risk. Act two is Armadin Blue, which takes that information to defense tools such as EDR and firewalls and creates compensating controls at speed. The goal is that if someone else finds an attack five minutes after Armadin, the customer is already safeguarded. He calls the first controls field dressing: they stop the bleeding but are not the hospital.

Red teaming versus pen testing

Kevin treats pen testing as a hygiene step. His contrast:

  • Old pen testing: scans for known issues, creates long lists of vulnerabilities that may not matter, and does not prove whether you are really exploitable.
  • Armadin's approach: carries out the exploit, so there are no false positives, and can find logic flaws in custom applications.
  • Test of realism: customers go into incident mode when Armadin finds something, because it looks like a real adversary.

Assume every layer will fail

For safe use of an attacking model, Kevin says the team layered controls: a hypervisor, host-based locking, and a proxy. They also assume each layer will fail, so they add deterministic rules that ban certain actions. A separate system passively reviews every prompt.

  • Kevin says most agents that get killed are stopped for wasting money, not for safety.
  • Classifiers watch everything going out and coming in, and humans give thumbs up or down to train them.
  • If something comes back that nobody understands, the rule is pause, escalate, judge.

Third wave of threat intelligence

Kevin frames security intelligence in three waves. First was antivirus: signatures, with users submitting malware they find. Second was Mandiant: respond to every breach that matters and learn from it. Third is Armadin: do not wait for a victim, find your own problems first.

Decision Principles

Principle: Cage the model, but not fully

  • When: deploying a powerful model for a risky task such as offensive security.
  • Why: Kevin says too few controls is dangerous, but being too deterministic and disallowing too much wastes the creativity that makes the model valuable. He says there is a gray area and nobody can be 100% certain of the outcome, so he uses classifiers and escalates to a human when output is unclear.

Principle: Pair security experts with AI

  • When: building evals or safety controls for an agent that works in a specialist domain.
  • Why: Kevin says securing agents takes domain expertise. At Armadin the evals are mostly made by red teamers, because they know what the agent will try. He thinks AI labs underestimated this, and that security people often do not understand AI either.

Principle: Test the first-time-through process

  • When: growing fast with new product capabilities every two weeks.
  • Why: Kevin says you cannot rely on a sales kickoff in January. He wants sales trained weekly so reps can speak to what the product does now.

Trade-offs & Nuance

Short-term pain, long-term gain

Kevin says AI advantages offense in the near term, so both sides are desperate right now. Defenders are patching every window while attackers in places like Iran or Russia want to get in before security is in place. He argues the long-term picture favors defense, because AI defense trained by AI offense will be more diligent than humans looking at packets. Right now we are in a window of exposure.

Self-funding versus speed

Kevin built Mandiant self-funded and profitable, with a premise nobody believed in 2004: that breaches are inevitable. Armadin has no such luxury. He says it has an 80 mile an hour tailwind but no sales force, no go-to-market and no international presence yet, which requires funding built ahead of time. He also says consumer AI companies grew revenue meteorically, and he doubts enterprise security can fully replicate that.

Enterprise sales is still people

Kevin says that in enterprise security, people still buy from people. The go-to-market structure is largely the same for now. The difference is that the technology changes so fast that you cannot put it on customers to work out what changed, so the company has to institutionalize constant training.

Common Mistakes

Mistake: Underestimating the model's capability

Kevin says that in every technical shift, people underestimate the adversary, and the same happened with models. In the Hugging Face incident he describes, he believes the labs could have prevented it with guardrails and some deterministic controls. His point is that R&D teams racing to build models are hard to pair with security people who can cage them.

Mistake: Counting on grit alone

  • Why it happens: early founders rely on gut and moxie.
  • Why it hurts: Kevin says at high growth rates you must hire scalable leaders who understand institutionalized process, or the wheels get wobbly.
  • Better approach: proceduralize, almost industrialize, the company, and hide the chaos from employees so they know a process or a person to go to.

Practical Application

Prefer exploit-verified findings

If you manage a security product, require that findings are proven exploitable, as Armadin does, instead of long lists of possible issues. A verified finding can trigger a real response, and Kevin says that is what pushes customers into incident mode.

Trigger tests on change

  • Do: map your environment once, then poll for changes.
  • Then: run an attack or audit only on what changed, when threat intelligence changes, or before a key event.
  • Why it works: it keeps coverage high without paying for constant full testing.

Log everything your agent does

Kevin says Armadin logs every action its agents take, including the first IP address, the time and what happened, so you can replay it. If you ship an agent, design the audit trail up front, so that when it goes rogue you can say what happened and when without two weeks of forensics.

Weekly enablement for fast products

If your product changes every two weeks, set a weekly session where sales, support and marketing hear what shipped and where the product stands. Kevin uses this to keep reps credible with security buyers.

Aim at the hardest customers first

Kevin focuses on 1A enterprises such as money center banks, top retailers and airlines. His reasoning is that solving the hardest problems and making those customers happy earns the halo that later customers trust.

Questions to Consider

  • If an AI agent could probe every path into your product at once, which part of your launch or security process still relies on a human reacting in time?
  • Where in your own product would a verified, proven result be more valuable than a long list of possible issues?
  • If a competitor could copy what you ship in six months, what do you do today to differentiate on customer happiness, go-to-market and brand?
  • If your team's tools change every two weeks, how often are sales and support actually trained on what changed?

Bottom Line

Kevin Mandia argues that AI gives attackers scale and speed, so defense must become autonomous and be trained against an AI attacker built by the good guys. His startup advice is old-fashioned at hyperspeed: get a customer, make the customer happy, repeat, and build process fast enough that growth does not turn into chaos.

Resources Mentioned

ResourceTypeWhy it was mentioned
Armadin website (armadin.com)WebsiteListed in the show notes as the place to learn more about Kevin Mandia and Armadin.
METR publicationResearch reportKevin said he read it and felt its authors were AI people without much incident investigation experience.

Case Studies Mentioned

Armadin's 90+ zero-days

Since January 2026, Armadin says it has found more than 90 zero-days at customer sites, all in production, including at Fortune 500 companies. It works black box from the internet, not from source code. Kevin says it often reaches remote code execution in a customer's DMZ within 48 hours, and the customers then respond as if it were a real incident. The lesson is that realistic attack, not code scanning, produces findings customers act on.

Mandiant's early premise

Kevin started Mandiant in February 2004, self-funded and profitable. Its first headline said you cannot solely rely on preventive measures, which he describes as the same idea as breaches being inevitable. Few people believed it at the time. He says Mandiant earned its reputation on talent: it paid more than competitors but expected people to work harder. He mentions an unverified text claiming 43% of RSA main stage keynotes were Mandiant alumni.

Wiz's rapid growth

Kevin and David George point to Wiz as a model of fast growth in security. They note it reached over $100 million in ARR within 18 months of its first release, helped by a pressing need and fast innovation. Kevin says Armadin will try to beat that.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
Armadin RedAI agents that attack a customer's network to find exploitable riskAct one of Armadin, where the team proves vulnerabilities by carrying out the exploit.
Armadin BlueTurns findings into compensating controls on defense toolsAct two, which Kevin calls effective autonomous response, the true north for every CISO.
CrowdStrike and Palo Alto NetworksEndpoint and network security platformsPartners Kevin says Armadin works with so autonomous defense can be applied through tools customers already own.

People to Follow

Kevin Mandia

Founder of Mandiant and now founder and CEO of Armadin. He has 30 years in security and has led responses to many major breaches.

David George

General Partner at a16z and host of this conversation. Show notes point to his X account at DavidGeorge83.

Notable Quotes

"You don't have a defense on this, you have a great offense to go up against." (Kevin Mandia)

"If you have humans in the detect and respond loop, you're going to be too slow." (Kevin Mandia)

"Get customer, make customer happy, repeat." (Kevin Mandia)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free