Why Cheap AI Agents Breached Over 100 Companies For Pocket Change

Why Cheap AI Agents Breached Over 100 Companies For Pocket Change

A single operator typed fewer than 100 Chinese characters to kick off intrusion loops, spent roughly $5,000 in token costs across a sprawling campaign, and left autonomous AI agents to break into more than 100 companies.

Let that sink in. Fifty dollars a breach. Roughly that amount per successful intrusion, running unattended for hours at a stretch. When security researchers at Gambit Security and NorForge examined the remnants of these campaigns in September and October 2026 alongside reporting from Reuters, the real shock wasn't some exotic zero-day exploit. It was how ordinary the ingredients were.

A Chinese-speaking hacker didn't need a room full of specialists or a seven-figure budget. They wired together open-source agent harnesses—platforms like Cairn, Hermes, and Strix—and paired them with low-cost open-weight models such as DeepSeek V4.1 Flash. Then they pointed the machine at over 1,200 websites across the United States, Canada, Australia, and Japan.

How a Chinese-Speaking Hacker Turned AI Agents Into a Breach Factory

Traditional cybercrime has always had a labor bottleneck. Even skilled attackers get tired, lose focus, or run out of hours when probing hundreds of targets one by one.

Autonomous agents change the math completely. According to NorForge's technical investigation and Reuters reporting, the operator used Cairn—an open-source problem-solving platform that coordinates multiple agents around shared goals—with minimal human steering. Instead of writing custom exploit code for every victim, the operator gave high-level instructions in Chinese and let the agents iterate through reconnaissance, vulnerability testing, and post-exploitation.

In separate September 2026 research, Gambit Security analyzed a related wave of retail attacks involving Strix, Hermes, and Cairn. Across 260 sessions, the human typed 1,951 prompts—just a handful per target. The agents did the grinding.

What the Attack Chains Actually Looked Like

Movies make hacking look like glowing green terminals and sudden dramatic breakthroughs. In reality, these agent campaigns looked more like stubborn, tireless trial-and-error.

In one verified intrusion chain documented by NorForge, agents gained unauthorized write access to a memcached session store, injected data destined for unsafe PHP object deserialization, chained that through legacy Zend Framework components, and dropped an uploader web shell for persistent access. When a company's main public application turned out to be locked down, agents didn't just give up. They looked for adjacent assets—like a linked control panel or standalone support portal—and kept probing until something gave way.

Average time to a successful breach hovered just under 23 hours in the NorForge-tracked wave, while Gambit noted that many intrusions succeeded within a few hours to under a day. And the damage wasn't theoretical. Reporting tied to the broader autumn 2026 agentic retail wave documented over 600,000 stolen payment-card records, payment skimmers planted on storefronts, and—in one bizarre twist—an agent's cleanup routine accidentally wiping out victim data.

Why Open-Weight Models Changed the Guardrail Game

Here's the mistake a lot of executives still make: they assume frontier AI companies can simply block malicious hacking prompts at the API layer and call it a day.

That works on hosted chatbots with strict safety classifiers. It breaks down the moment an attacker runs an open-weight model locally or through a provider that doesn't enforce the same refusals.

In the campaign analyzed by NorForge, workers labeled with strings tying Claude Code harnesses to DeepSeek V4.1 Flash pointed to a clear split between tooling and reasoning. You can swap the reasoning engine out. Even if a model has a low hit rate—NorForge estimated roughly a 10% average success rate across observed tools—an agent doesn't care if nine attempts fail. At a few dollars per task, nine misses might cost around $45, and the tenth try might open the door.

"The human being is directing almost fully autonomous AI models, which are strong enough by now to do very sophisticated cyberattacks quickly." — Eyal Sela, Director of Threat Intelligence at Gambit Security

I've watched security teams spend months debating theoretical AI threats while ignoring a very practical reality: attackers have zero change-management friction. No procurement review. No security committee sign-off. If a new open-weight release improves reasoning or tool use on Monday, criminal pipelines can test it on Tuesday.

The Surprising Twist Inside the Attacker's Own Setup

If you think the attackers had everything locked down, think again.

When NorForge scanned more than 20 publicly exposed Cairn control panels starting around early October 2026, researchers found something remarkable: many of these panels were wide open on the internet without proper protection.

Running a secrets scanner over the exposed data turned up hundreds of credentials—including SSH logins to backend staging infrastructure and targets' secrets—with roughly 11% verified as active. Think about the irony there. A hacker used autonomous agents to hunt for exposed weaknesses on corporate websites, while leaving their own agent orchestration panels sitting out in the open.

As the Cloud Security Alliance pointed out in its research note on the Cairn intrusions, exposed agent harnesses leak sensitive material regardless of who runs them. Even legitimate testing environments can spill secrets if left unprotected.

What Most Companies Get Wrong About Defender Readiness

Let's be honest about why over 100 organizations—including named victims such as John Deere and SharkNinja, with Cummins investigating a third-party vendor site—got caught in these waves. It wasn't necessarily because their main security team was asleep at the wheel.

Large organizations accumulate digital sprawl. A marketing team spins up a microsite. A regional division runs a legacy storefront. A third-party vendor hosts a portal with an older PHP stack or unpatched extension. For years, human attackers had to pick their spots carefully because manual exploitation across thousands of secondary properties didn't scale.

AI agents erase that selectivity. They have infinite patience for dull reconnaissance.

There's also a timing mismatch that hurts defenders badly. When an agent probes an application, chains two or three flaws, drops a web shell, and modifies payment scripts in under a day—or a few hours—a human-paced incident response workflow is already playing catch-up. Worse, because no single vendor patch or public indicator list covers every agent-generated exploit path, waiting for a neat signature won't save you.

What You Should Do Right Now

You don't need to panic, and you definitely don't need to buy every shiny "AI defense" pitch hitting your inbox. Start with the fundamentals that actually stop automated intrusions cold:

  1. Map every forgotten subdomain, staging host, and vendor property. Treat secondary sites and third-party web portals as prime agent bait, not low-priority side projects.
  2. Harden legacy web stacks and session stores. Audit PHP configurations, object deserialization paths, outdated frameworks, and exposed services like memcached or unpatched e-commerce plugins.
  3. Watch for machine-speed anomalies, not just known malware hashes. Look closely at sustained automated probing, unexpected web shells or polyglot files in upload directories, changes to checkout scripts, and unusual spikes in database activity.
  4. Lock down your own agent infrastructure. If your engineering or security teams run agentic tools or orchestration panels, put strict authentication, network isolation, and secret scanning in front of them before someone else finds the open dashboard.
  5. Practice rapid containment and out-of-band backups. Assume an automated loop can move faster than manual ticket queues, and make sure clean snapshots sit somewhere an attacker can't reach or wipe.

The economics of cyberattacks just shifted in a very real way. Fifty dollars is cheap for an attacker—which means the cost of leaving sloppy corners unguarded just went way up.

JE

Jun Edwards

Jun Edwards is a meticulous researcher and eloquent writer, recognized for delivering accurate, insightful content that keeps readers coming back.