/
세분화

AI-Powered Cyberattacks: Why Friction Trips Up Machine-Speed Attackers

Before she moved to Wall Street, Rachel Wilson led cyber exploitation operations at the U.S. National Security Agency. She remembers a room full of about 300 security analysts who spent all year combing through source code for zero-day vulnerabilities.

In a good year, they found five. Today, frontier AI models surface flaws by the hundreds and thousands.

The cost of exploiting those flaws has collapsed, too. Attacks that once took hundreds of human hours now cost about $5, by the estimates her team shares with its board.

Wilson, now chief data officer at Morgan Stanley Wealth Management, told that story on a panel at the Illumio World Tour stop in New York in October. She was joined by Illumio CEO Andrew Rubin; Fred Rica, who leads the cyber practice at BDO; and Sherrod DeGrippo, vice president of threat intelligence at Palo Alto Networks Unit 42.

The panel spent 40 minutes on what happens to cyber resilience when AI-powered cyberattacks move faster than any human team can respond. Their answer kept coming back to the thing security teams have spent years trying to remove: friction. Every boundary, checkpoint, and approval that slows an attacker gives defenders time to act. And time is what machine-speed attacks take away.

AI-powered cyberattacks have broken the old math

For 40 or 50 years, cybersecurity has been a human game, Rubin said. Human attackers faced human defenders, and both sides worked at human speed with the tools they had.

“The math no longer holds water,” he said. “If you give attackers models and the attack now moves at machine speed, your defense posture at human speed no longer works.”

Wilson’s numbers show what that looks like in practice. Firms that spend hundreds of millions of dollars a year on defense now face attackers with little skill, almost no budget, and AI tools that do much of the work for them.

DeGrippo compared the shift to trading a hammer for a pneumatic nail gun. A hammer drives one nail at a time, and a slip usually costs you a sore thumb. A nail gun fires with one squeeze of the trigger, and a slip can send a nail straight through your hand.

“We’ve put nail guns in everyone’s hands on the planet,” she said.

In her view, script kiddies, the low-skill hackers who stitch together pre-made tools, “are the most powerful people on the planet now.” Rica added that AI is making attacks faster, cheaper, more scalable, and more persistent.

“Time is the battlefield now,” he said.

Why security friction has a bad reputation

If time is the battlefield, friction is how defenders win it back. That’s a hard sell, though, because friction has never been popular.

After about a decade of adoption, segmentation still has “a fairly crusty and at times difficult reputation” because it adds friction, Rubin said. A flat network with no firewalls, subnets, zones, or VLANs would be fast and easy to use. It would also be “wildly, wildly insecure,” he said, and an attacker would own it in seconds.

He pointed to the most secure building he’s ever visited, the White House. It’s hard to get in, hard to move around, and even hard to leave. All that friction gives the people running it a clear view of everything that happens inside.

“Friction equals security,” he said.

The trouble is that many companies have spent years removing friction. They’ve worked to keep developers happy and make the cloud easy to use.

Rubin described a recent visit with a CISO at a global systemically important bank. The bank had moved from zero to 30,000 developers on an AI coding tool in 90 days, procurement included. Yet a firewall refresh or a segmentation project still feels like a multi-year effort to most security teams.

AI adoption is moving at machine speed while security is still moving at human speed.

Segmentation buys back the time AI attacks take away

Rubin described segmentation as a way to “buy yourself time.” When you segment a network, lateral movement gets much harder for attackers. They also get caught more easily, because they have to cross boundaries that policy says they shouldn’t.

Every one of those boundaries slows them down and gives your team time to investigate before a small problem turns into a big one.

Wilson said banks have always needed segmentation for business reasons. Doing business in China required it. So did having clients and employees in the same building, or keeping a buy side and a sell side apart. Acquisitions are another. Strong network segmentation lets a buyer bring new networks on board while keeping the risk contained.

Now, she said, “the number of cases that we need to think about this has increased astronomically.”

She also drew a clear line between two kinds of companies. Those with extensive, real-time visibility already in place will have a huge advantage when they need to apply segmentation fast. Those that have run flat networks all along will “find this an uphill battle,” she said.

For security engineers, that’s a strong case for mapping traffic and applying microsegmentation now, before an AI-driven attack forces the issue.

Human checkpoints and AI agent governance are friction, too

Rica made the case that friction goes well beyond the network. A human in the loop is friction. So is an AI agent that has to ask “a grown-up,” in his words, before it takes a consequential action.

He tells clients to give employees permission to pause and be a little skeptical. Is that really the CEO on Zoom, or is it a deepfake? Should this payment really go to this vendor?

His biggest worry is rogue AI agents. Say you ask an agent to book you an exercise class, and it hacks the waiting list to get you in. The agent did exactly what you told it to do, using access you didn’t realize it had.

That’s why agent governance needs the same rigor you’d apply to any other identity. His checklist is a good place to start:

  • Who owns the agent?
  • What does it do, and what can it reach?
  • How long should it live, and when does it get recertified?
  • Where’s the kill switch if it goes off the road?

In his view, AI adoption is outpacing AI governance and exposing every governance gap that was already there.

Wilson’s team is adding friction to how it builds software, too. The old model of writing code, shipping it, and running a pen test three months later is over, she said. AI-driven testing will happen well before release, and new code will roll out in stages.

“This is going to slow things down. It’s going to increase friction. But it’ll also give us much more confidence that the code we ship is secure,” she said.

Cyber resilience starts with breach containment

Resilience is, by definition, the ability to recover from a problem, Rubin said. You don’t need it if you never have problems.

AI has made that ability a board-level conversation, and it’s quickly becoming a regulatory one, too.

Security teams have long assumed a breach is only a matter of time. The whole panel agreed the focus now belongs on containment.

“It’s the idea that we’re never going to be perfect, but that we’re going to minimize the blast radius. And our boards of directors have to accept that is enough,” Wilson said.

Rica also linked resilience to dependency. After 20 years of trading resilience for efficiency, many firms rely on single sources for critical services. Knowing those dependencies, he said, can make “the difference between it becoming an incident and an existential event.”

Your head start on AI-powered cyberattacks won’t last

Defenders still have a head start, the panel agreed. Today’s most powerful frontier models aren’t widely available. But distilled and open-source versions are on the way. And Wilson warned that companies outside the Fortune 50 fall below “the cybersecurity poverty line real quick.”

So add friction now. Segment your network before an attacker maps it for you, put a human checkpoint in front of risky actions, and give every AI agent an owner and a kill switch. Each layer buys time, and against AI-powered cyberattacks, time is what keeps a small breach from becoming a catastrophic one.

Rubin said the right outlook is to be an “optimistic realist.” The realist knows that what once took 300 analysts a year now takes a model and a few dollars. The optimist knows what to do about it: slow the attacker down.

The best time to find your attack paths is before an AI-powered attacker does. Try Illumio Insights free to map how your workloads connect and where an attacker could move. Then download The Mythos Mandate to learn how breach containment helps you stay resilient at machine speed.

관련 문서

지금 일루미오 인사이트 체험하기

AI 기반 관찰 가능성을 통해 위협을 더 빠르게 탐지, 이해, 차단하는 방법을 알아보세요.