Skip to main content
Return to TrendAI™ 資安部落格
AI & emerging technologies APTs

AI 駭客不再單打獨鬥:從入侵百家企業到國家級網路攻擊的新威脅

AI 協助單一攻擊者以每次約 50 美元的成本,入侵超過 100 家企業。然而,更大的風險在於:這些存取權限接下來可能落入誰的手中?

Perspectives AI governance Cyber espionage

Key Takeaways

  • AI 大幅降低網路攻擊門檻: 單一攻擊者利用 AI 代理程式(AI Agent),對超過 1,200 個網站發動入侵嘗試,成功入侵逾 100 家企業,平均每次成功入侵的成本僅約 50 美元。這顯示 AI 已大幅降低網路攻擊的技術與成本門檻,使個人攻擊者也能發動大規模攻擊。
  • 單人駭客可能成為國家級網路攻擊的助力: 個人攻擊者取得的系統存取權限,可能透過地下犯罪網路或地緣政治關係(如 Pax Mafiosa 模式),轉移至國家級威脅行為者手中,進一步將遭入侵企業的系統變成攻擊客戶、合作夥伴及供應鏈的跳板。
  • 企業資安防禦須從單次事件處理轉向持續監控: 資安團隊不應將單次入侵視為孤立事件,而應以存取權限可能已遭轉移或即將被轉手的情境進行風險評估與事件應變規劃,持續監控內部 AI 代理程式行為、追查命令與控制(C2)通道,並盤點第三方合作夥伴管理的對外曝險資產。

It took a prompt of fewer than 100 Chinese characters for a single cyber attacker to break into websites at more than 100 companies, including Deere, SharkNinja, and Cummins.

The attacker's agents attempted intrusions against more than 1,200 sites. They spent roughly 23 hours probing each successful target, and the entire attack cost around $50 per breach.

The headline is that one person with a few thousand dollars can now run a campaign that used to take a crew. But the bigger story is what happens to that person next.

Lone wolves don't stay lone for long

In Russia, a lone cyber actor who accumulates this kind of access does not remain independent for long. Whether or not they ever meant to serve the state, the intelligence services find them and harness what they have gathered for the nation state's purposes.

China runs a different model, built on contractors and front companies rather than a protection racket, but the result is converging: the state collects what its lone attackers collect.

In Russia, it's an arrangement I call the Pax Mafiosa.

The Pax Mafiosa took hold in earnest around 2013 , and it runs on three implicit rules:

  1. You don't hack anything that uses your country's language settings.
  2. You share access to anything you have, especially if it's sensitive and you think it could be useful.
  3. And when called upon to be patriotic, you go after the specific targets you are given.

In exchange, the criminal is left alone. Their compensation is untouchable status. That is why the major cybercrime cartels operate with impunity from Western law enforcement.

Notice that none of the three rules requires skill. The second requires only inventory: the same access criminals sell to each other on underground forums is provided freely to the intelligence services in order to enjoy the protection racket.

AI just expanded the inventory

For two decades, the constraint on this model was talent. A state could only harvest as much access as its skilled underground could gather, and producing skilled operators took years. AI agents remove that constraint. The attacker in the Reuters story did not find a hidden admin panel or spot a subtle input-handling flaw. They supplied a prompt and a budget.

That is the point. The state does not need this person's skills. It needs what he collected: more than 100 footholds in American, Canadian, Australian, and Japanese companies, held by someone within reach of rule two. The agent did the technical work, but the state gets the operator and the footholds. And it still wants the operator, because a criminal cutout provides the deniability a government-run agent never can.

It is also worth noticing what this attacker did not touch. A Chinese-speaking actor hitting American, Canadian, Australian, and Japanese sites and no Chinese ones looks a lot like rule one in practice, though language is not proof of location. Meanwhile in Russia, jailbroken AI forums are where these actors gather, and we already see North Korean crews riding command-and-control infrastructure that Russian cartels built. The lone wolf of October becomes the auxiliary of December.

This is why I am less worried about the AI "dragon" breaking out of its box than about who climbs on its back. AI breaking loose is a theoretical risk. State-aligned actors riding it is happening now.

What this means for defenders

Three things change for defenders.

First, the amateur breach is no longer a low-severity event. The access a lone operator gains today can be in a state actor's hands tomorrow under the rules of the Pax Mafiosa. A compromised product-support site or a vendor's web app is a foothold with a future owner.

Second, the goal is no longer theft and exit. The adversary of today wants to hijack your digital transformation and use it to attack your customers, your partners, and the agencies you serve. Worst case is not ransomware. Worst case is your infrastructure becoming the island hop.

Third, the kill chain is compressed. Recon, tunneling, backdoors, disabling endpoint protection, privilege escalation, credential harvesting: all of it ran autonomously. What took an adversary days now takes an agent minutes.

Traditional cyber hygiene alone will not solve this. Assuming persistence and watching behavior continuously will.

Five things to do before the adversary's next owner arrives

“One must perceive that which one cannot see.” - The Book of Five Rings

  1. Treat every intrusion as a handoff in progress. Assume the access has already been shared or will be. Scope the incident for what a nation state would do with that foothold, not what the original intruder did.
  2. Do not terminate command-and-control the moment you find it. Killing the first C2 tells a sophisticated adversary you are onto them and invites a destructive response. The second and third C2s will be on sleep cycles or hidden in image files. Map the attack path, find them all, then cut them at once.
  3. Govern your own agents as if you had a known criminal on the payroll. Least-privilege tool control, isolated context, independent output validation. For each agent, ask: what can it reach, what would misuse look like, and would you see it?
  4. Monitor agent behavior continuously, east-west as well as north-south. This campaign was caught only because a researcher was scanning for exposed agent consoles. Continuous behavioral anomaly detection for agents is the AI-era equivalent of EDR.
  5. Hunt for the third-party foothold. Two of the three named victims were hit through a standalone site or a vendor-operated web app. Inventory every external asset your partners run on your behalf and make sure your threat hunting covers them. That is where the island hop starts.

Plan for the second owner

We don't yet know whether the access from this campaign has been sold, shared, or simply abandoned. But the economics now favor the first two. A $50 breach by an amateur is the cheapest reconnaissance a hostile state has ever been offered, and the rules of the Pax Mafiosa give the amateur every reason to hand it over. Defenders who plan for that second owner, rather than the first intruder, are the ones most likely to see it coming.

Join TrendAI™ at AWS re:Invent 2026

Tom will build on this idea at AWS re:Invent alongside CISO, author, and podcast host Ashish Rajan in their breakout session:

Threat research: How attackers use AI to persist and how to evict them (SEC380-S)
Thu., Dec. 3 at 11:30 a.m. PT
The Venetian, Level 2, Ballroom F, Content Hub, Red Theater