The Hacker Who Built an Army of Robots: How AI Agents Are Turning Cyberattacks Into Automated Operations

The fortress had survived countless attacks.
Its walls were reinforced. Its gates were guarded. Its defenders regularly inspected damaged structures and replaced vulnerable locks.
Every time a weakness was discovered, engineers repaired it.
The fortress seemed prepared.
But in a laboratory hidden beyond the surrounding mountains, a criminal scientist was working on something extraordinary.
He had spent years attacking fortresses himself.
He knew how to examine their defenses, identify weaknesses and search for entrances.
But every operation required time.
He could investigate only so many targets.
He could test only so many doors.
And whenever a fortress repaired its vulnerabilities, he had to study the changes before attempting another attack.
Then he built his first robot.
It could inspect a fortress without becoming tired.
His second robot could analyze engineering documents.
A third could compare old and new locks.
A fourth could test whether a discovered weakness remained exploitable.
Soon, the laboratory was filled with machines.
The scientist gave them their objectives.
The robots began working.
And the criminal discovered that he no longer needed to personally perform every step.
He had transformed hacking from a series of manual operations into a coordinated production line.
On November 16, 2026, Google’s Threat Intelligence Group (GTIG) reported that cybercriminals and suspected espionage actors were increasingly using autonomous AI frameworks to accelerate cyberattacks.
The research described offensive frameworks, automated vulnerability analysis, exploitation of known flaws, stolen AI infrastructure and phishing platforms enhanced with large language models.
The underlying attacks were not necessarily new.
What was changing was the speed at which they could be developed, coordinated and executed.
Phase 1 — The Criminal Scientist’s Laboratory
The scientist’s first challenge was building machines capable of performing different tasks.
A single robot could inspect a door.
But an entire army could examine multiple entrances, share findings and carry out more complex operations.
In cybersecurity, these machines correspond to AI agents.
Unlike a conventional chatbot that simply answers a question, an agentic system can use tools, execute defined tasks, interpret results and determine subsequent actions within its configured permissions.
Multiple agents can be organized into a coordinated workflow.
One may analyze a target.
Another may inspect code.
A third may evaluate potential vulnerabilities.
Another may help develop or test an exploit.
Google reported adversarial use of several open-source frameworks, including:
ARTEX
CAIRN
CyberStrikeAI
PentAGI
These tools and related frameworks can also have legitimate uses in authorized security testing.
Their existence is not inherently malicious.
The concern arises when attackers deploy them against systems without permission.
GTIG observed financially motivated threat actors and suspected China-aligned espionage actors experimenting with or using agentic offensive capabilities.
The robots were not dangerous merely because they existed. They became dangerous when the scientist gave them unauthorized targets.
Phase 2 — The Robots Learn to Work Together
The scientist assigned his machines different responsibilities.
One searched the fortress perimeter.
Another examined the walls.
A third investigated access points.
A fourth evaluated the findings and prepared the next action.
The scientist could supervise the operation without personally carrying out every task.
This illustrates the significance of multi-agent offensive frameworks.
Traditional attacks often require a human operator to move repeatedly between tools, analyze outputs and decide what to do next.
Agentic systems can automate portions of that process.
Google described frameworks being used to support discrete phases of real intrusions, rather than merely producing theoretical attack suggestions.
The distinction matters.
An AI model generating a possible command is different from a framework connected to tools that can execute commands, inspect their results and continue the workflow.
That ability can reduce the amount of manual intervention required.
However, autonomy varies between frameworks and operations.
The research does not establish that every observed intrusion was conducted entirely without human involvement.
The scientist still selected the fortress. But increasingly, his machines could decide how to approach its defenses.
Phase 3 — The Fortress Repairs Its Walls
One morning, the fortress engineers discovered a weakness.
They repaired it immediately.
A damaged gate was replaced.
A vulnerable locking mechanism was redesigned.
The engineers published documentation explaining which components had changed.
The defenders considered the repair a success.
But the scientist saw something else.
The new engineering plans revealed which parts of the old design had required attention.
He handed the old and new plans to his robots.
Their assignment was simple.
Find the difference.
In cybersecurity, this resembles patch differential analysis.
Researchers compare vulnerable and patched versions of software to understand what changed.
The technique has legitimate defensive applications.
It helps security researchers understand vulnerabilities and evaluate whether fixes are effective.
But attackers can use the same information to identify weaknesses in systems that have not yet been updated.
Google warned that threat actors are using AI-assisted workflows to accelerate this process.
Rather than researching an entire application from scratch, they can focus on a smaller collection of materials:
Source-code differences between software versions.
Release notes.
Security advisories.
Public vulnerability research.
Existing proof-of-concept code.
The fortress had repaired its weakest gate. But the repair plans could help the robots identify exactly where the old gate had been vulnerable.
Phase 4 — The Robots Turn Repairs Into Attack Plans
Once the robots identified a suspicious change, they began examining its implications.
Why had this particular function been modified?
What condition had the developers corrected?
Could an older version still be vulnerable?
Could the weakness be reproduced?
This is the process of turning vulnerability information into a potential exploit.
Google highlighted the use of AI-driven pipelines to accelerate the analysis and weaponization of previously disclosed vulnerabilities.
These are often called n-day vulnerabilities: security flaws that are already publicly known, typically with patches available, but that may remain exploitable in unpatched systems.
The article cited activity involving PaperCut NG/MF vulnerabilities in September 2026 as an example of accelerated exploitation-related research.
The operational concern is the shrinking interval between disclosure and weaponization.
Historically, attackers might need substantial manual effort to study a patch and produce a working exploit.
AI-assisted analysis can reduce some of that effort.
That does not mean every patch can automatically be transformed into a reliable exploit.
The outcome still depends on the vulnerability, the available information, the target environment and the capabilities of the tools involved.
Nevertheless, the potential advantage is significant.
The defenders were racing to repair every vulnerable gate. The robots were racing to understand the repairs before every fortress had installed them.
Phase 5 — The Army Moves at Machine Speed
The scientist had once needed to perform every investigation sequentially.
Now multiple machines could work on different tasks.
Some examined software.
Others investigated exposed services.
Others tested potential attack paths.
The machines could repeat tasks and process information at a pace difficult for a human operator to match.
Google characterized this as an operational shift toward attacks conducted at machine speed.
The underlying techniques remained recognizable.
Attackers still searched for vulnerabilities.
They still attempted to obtain access.
They still exploited credentials, exposed services and misconfigurations.
But automation could make those activities faster and easier to scale.
This creates a challenge for defenders relying on slow operational cycles.
A vulnerability discovered during a weekly scan might already have been analyzed and targeted by an automated adversary.
A patch scheduled for installation several weeks later could leave a significant exposure window.
The robots had not discovered a magical weapon. They had made the scientist’s existing weapons faster to manufacture and easier to deploy.
Phase 6 — The Scientist Discovers More Powerful Machines
The laboratory continued expanding.
Some robots were designed for autonomous security testing.
Others could generate tools or execute code.
Still others could help the scientist select and route tasks between different AI models.
Google identified additional technologies being used or explored in offensive workflows.
One was Agent Zero, an open-source agent framework featuring autonomous code execution and dynamic tool generation.
Another was ZeroHollow, described as an inference service used for exploit development and routing requests to language models.
These technologies illustrate a broader ecosystem.
The offensive capability does not necessarily reside in one AI model.
It can emerge from the combination of models, orchestration frameworks, tools, external services and the permissions granted to them.
That combination can make a system more operationally capable than a standalone chatbot.
It can also introduce risks when deployed without adequate authorization, monitoring or restrictions.
The scientist’s greatest invention was not a single powerful robot. It was the machinery that allowed different robots and tools to work together.
Phase 7 — The Robots Discover the AI Power Station
The scientist’s army needed computing resources.
AI models and agentic workflows can consume substantial processing capacity.
Access to cloud-hosted models and inference services can also require valuable credentials.
The scientist realized that stealing access to an AI power station could be profitable.
Google described an activity known as LLMjacking.
In these attacks, criminals obtain credentials or cloud resources that provide access to large language model services.
Those resources can then be abused without authorization.
Google also reported an underground market in which stolen AI access was resold through proxy services, sometimes described as LLM mixers.
The article named APIMart and JetNight as examples.
The financial attraction is straightforward.
AI infrastructure costs money.
If criminals can use another organization’s credentials or computing resources, they may avoid paying those costs themselves.
They may also resell access to other customers.
The consequences for victims can include unauthorized consumption, unexpected costs and exposure of compromised cloud resources.
The article does not establish that every stolen credential was used to launch an autonomous intrusion.
LLMjacking is a related form of AI-enabled cybercrime, not necessarily a required stage in the agentic attacks.
The scientist had discovered that he could do more than build robots. He could steal the electricity needed to operate them and sell access to the power station.
Phase 8 — The Robot That Could Design Its Own Disguises
Not every machine in the laboratory was designed to break through technical defenses.
One specialized in deception.
It could prepare convincing messages.
It could adapt a fraudulent website.
It could modify the language of a phishing campaign.
And it could do so from ordinary written instructions.
Google highlighted Jieyou, a Chinese-language phishing-as-a-service platform that integrated an LLM-powered chatbot assistant.
The assistant allowed customers to configure phishing panels and modify templates through natural-language commands.
This lowered the amount of technical expertise required to customize certain phishing operations.
Instead of manually editing every component of a malicious page, a customer could describe the desired changes.
The system could assist with implementing them.
The result is another example of AI acting as a force multiplier.
A less-experienced criminal may gain access to capabilities that previously required greater technical knowledge.
However, the article does not demonstrate that Jieyou’s chatbot independently conducted every stage of a phishing attack.
Its reported role was to help configure and customize the phishing infrastructure.
Not every robot carried a weapon. Some manufactured convincing disguises for the scientist’s other operations.
Phase 9 — The Army’s Greatest Advantage Wasn’t Intelligence
The scientist watched his machines work.
Some succeeded.
Others failed.
Some produced useful findings.
Others followed unproductive paths.
But even unsuccessful attempts could be repeated or adjusted.
The machines didn’t need to be perfect to create a problem.
They needed to reduce the cost and time required to attempt an attack.
This is one of the most important implications of Google’s research.
AI may lower the barrier to entry for certain offensive tasks.
It can help attackers analyze technical information, connect existing tools and automate repetitive operations.
That can increase the number of targets an adversary is able to examine.
It may also allow attackers to adapt their workflows more quickly.
But the distinction between capability and demonstrated impact remains essential.
The existence of an AI-powered exploit-development pipeline does not prove that every generated exploit works.
Likewise, the use of an agentic framework does not mean every target was compromised.
The practical risk lies in the combination of accessible tools, reduced manual effort and existing security weaknesses.
The scientist didn’t need an army of flawless machines. He needed enough machines to keep searching until one found a vulnerable entrance.
Phase 10 — The Defenders Change Their Strategy
The fortress commanders recognized that their existing procedures were becoming too slow.
They had relied on scheduled inspections.
They had repaired weaknesses according to established maintenance cycles.
They had investigated suspicious activity after receiving alerts.
But an automated adversary could move through several stages of an attack before a manual response began.
The defenders needed to adapt.
Not by abandoning established cybersecurity principles.
But by applying them more quickly and consistently.
The most important measures included:
1. Accelerating vulnerability management
Prioritize security fixes for internet-facing systems and vulnerabilities with evidence of active exploitation.
2. Reducing exposed attack surfaces
Remove unnecessary public access to administrative interfaces, services and applications.
3. Strengthening identity security
Use least privilege, multifactor authentication where applicable and carefully scoped credentials.
4. Protecting AI infrastructure
Secure model API keys, cloud credentials, inference services and associated resources against unauthorized use.
5. Improving behavioral detection
Monitor for unusual reconnaissance, rapid sequences of suspicious activity, unexpected tool execution and abnormal access patterns.
6. Segmenting critical environments
Limit the ability of an attacker to move from an initially compromised system into sensitive infrastructure.
7. Automating defensive workflows
Use appropriate automation to reduce the time between detection, investigation and containment, with human oversight for consequential actions.
8. Monitoring third-party tools and frameworks
Maintain visibility into authorized AI agents, penetration-testing tools and their permissions.
9. Testing incident-response readiness
Prepare for attacks in which reconnaissance, exploitation and subsequent actions occur much faster than traditional manual operations.
10. Maintaining continuous exposure awareness
Identify vulnerable assets and misconfigurations before automated attackers can exploit them.
These recommendations reflect established security practices.
The new urgency comes from the possibility that attackers can execute familiar techniques with much greater speed.
The fortress didn’t need stronger walls alone. It needed defenders capable of recognizing and responding to threats before the robots reached the next gate.
Phase 11 — The Robots Are Tools, Not Independent Criminals
The scientist stood before his machines.
They could examine systems.
They could analyze vulnerabilities.
They could assist in exploit development.
And some could perform multiple tasks with limited supervision.
But they had not independently chosen to become criminals.
The malicious objectives came from the people directing or configuring them.
This distinction is essential when discussing AI-enabled attacks.
Agentic frameworks can be useful in legitimate penetration testing, vulnerability research and defensive automation.
The same capabilities can be misused by criminals.
Google’s findings concern that malicious use.
They do not establish that AI models have developed criminal intentions or that every autonomous security-testing system is dangerous.
Nor do they demonstrate that human attackers have disappeared from cyber operations.
Human operators still select targets, establish objectives, provide infrastructure and decide how to use the results.
What is changing is how much of the operational workload can be delegated.
The robots were not replacing the criminal scientist’s intentions. They were multiplying his ability to act on them.
Phase 12 — A New Race Between Attackers and Defenders
The implications extend beyond any individual framework.
As AI capabilities become more accessible, attackers may be able to automate increasingly complex sequences of activity.
That creates pressure on organizations to shorten their own defensive timelines.
A security patch may become an immediate research target.
A newly exposed service may attract automated reconnaissance.
A stolen cloud credential may be monetized through unauthorized AI access.
A phishing operator may customize deceptive pages without manually writing their code.
These are different threats.
But they share a common theme.
Automation reduces the amount of human effort required to perform operations that were previously slower and more labor-intensive.
Google’s research suggests that the competitive advantage in cybersecurity may increasingly depend on operational speed.
That does not mean defenders must surrender control to autonomous systems.
It means organizations should combine sound security architecture with timely remediation, monitoring, carefully governed automation and effective incident response.
The objective is not to win a contest over who can deploy the largest army of robots.
It is to make unauthorized access difficult, detect suspicious behavior quickly and prevent an initial weakness from becoming a major compromise.
The Final Scene — The Scientist Steps Away From the Controls
The fortress was quiet.
Its guards patrolled the walls.
Its engineers continued repairing vulnerabilities.
And its commanders believed they understood the enemy.
They expected a criminal searching for a weak entrance.
A hacker studying a vulnerable application.
An intruder attempting to exploit a damaged lock.
But beyond the mountains, the scientist’s laboratory was still operating.
One robot examined a newly released patch.
Another compared it with an older software version.
A third analyzed the difference.
Others continued searching for exposed systems and opportunities.
The scientist watched the machines from his control room.
He no longer needed to personally investigate every target.
He had created an operation in which parts of the attack could continue without constant manual direction.
That was the warning behind Google’s November 2026 findings.
Cybercriminals were beginning to use autonomous frameworks, AI-assisted exploit development and language-model-powered services to accelerate familiar malicious activities.
The underlying weaknesses remained recognizable.
Unpatched software.
Exposed systems.
Compromised credentials.
Poorly protected infrastructure.
But the speed of exploitation was changing.
The scientist switched off his monitor.
The laboratory lights dimmed.
And behind the glass, the machines continued working.
The most frightening thing about the hacker’s new army wasn’t that the robots had learned to break into fortresses.
It was that they no longer needed him to guide every step.
Source: The Hacker News — Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches (November 16, 2026). Additional context: Unbiased Headlines +1




Comentarios