Google Uncovers AI Agents Built To Hack On Their Own
Google Uncovers AI Agents Built To Hack On Their Own
Cyberattacks are becoming more automated, and Google says the threat has reached a worrying new level.
Criminal groups are no longer just asking AI chatbots to write suspicious emails or generate malicious code. Some are now using AI-powered agents that can investigate software weaknesses, develop ways to exploit them, and carry out multiple stages of an attack with little or no human involvement.
In a new threat intelligence report, Google's security researchers describe how attackers are building systems that can do much of the hacking work themselves.
The concerning part? These agents are not simply suggesting what a hacker should do next. They can take action, study the results, and adjust their approach as they go.
And this could change how quickly organizations need to respond to cyber threats.
This Is More Than AI Writing Malicious Code
For a while, one of the biggest cybersecurity concerns surrounding AI was its ability to help people write phishing messages, generate malicious scripts or understand technical information.
But the situation is evolving.
Google Threat Intelligence Group (GTIG), working with cybersecurity firm Mandiant, says it has observed attackers moving beyond ordinary AI assistance and using systems built around multiple autonomous agents.
These systems can divide a complicated operation into smaller tasks. One agent might search for weaknesses, another can examine the findings, while another attempts to develop and test a way to exploit the vulnerability.
They can then share information and coordinate their next moves.
Think of it as a digital team working together on a cyberattack, with AI agents handling different responsibilities instead of waiting for a human to give instructions at every step.
Google says this shift is helping attackers move through multiple stages of an intrusion at a speed that would be difficult to achieve through manual work alone.
Google Found an AI System That Could Hack Automatically
One of the most concerning discoveries in Google's report involved a system researchers called exploit_agent.
Google found an exposed staging server hosting this multi-agent framework, which could work with different advanced and locally hosted AI models.
Its job was to look for weaknesses in popular web applications.
Here is how it worked.
When software developers release a security update, they usually change the code responsible for fixing a vulnerability. By comparing the older version with the updated version, researchers can sometimes work out what the original weakness was.
This process is known as differential analysis.
The exploit_agent system was designed to automate much of this process.
It could:
Compare software versions before and after security updates.
Identify changes that might reveal a vulnerability.
Analyze potential weaknesses.
Generate code designed to exploit those weaknesses.
Automatically execute the generated exploits.
The important detail is that Google observed this process operating without a human needing to approve every step.
Researchers also discovered scripts associated with activities that could happen after an attacker gains access, including searching internal systems, attempting to guess credentials, and installing webshells.
A webshell is a malicious tool that can give an attacker remote access to a compromised server.
Based on activity found on the exposed server, Google said the operator appeared interested in websites in France, particularly in the media, entertainment, and education sectors.
This is a significant development because it shows how AI can help connect vulnerability discovery and exploitation into a largely automated process.
The AI Hacking Tools Attackers Are Using
Google's report also identified several open-source frameworks being used or tested by threat actors.
Some of the names include ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI.
These tools are not all identical, but they share an important idea: using AI models and specialized software tools to automate different parts of cybersecurity operations.
For example, CAIRN allows specialized agents to work together towards a goal. It uses a shared information space where agents can exchange findings and coordinate their actions.
CyberStrikeAI connects AI models with a large collection of security tools, scripts and attack procedures.
PentAGI is designed to make complex offensive security operations more accessible by automating activities such as reconnaissance and vulnerability scanning.
Google says it has observed CAIRN and PentAGI being deployed in real-world activity, while also seeing experimentation involving other frameworks.
The concern is not simply that these tools exist. Many cybersecurity tools can be used for legitimate security testing.
The problem is that criminal groups can use similar capabilities to investigate and attack systems without permission.
Google also reported evidence that a suspected China-linked espionage group was developing its own automated attack framework, although that does not mean every tool or incident identified in the report was connected to China.
Are These AI Agents Impossible to Stop?
Not exactly.
Google says the current generation of autonomous hacking frameworks still has weaknesses.
For example, some generate unusually large amounts of network traffic, leave recognizable traces, and produce activity that security teams can detect.
Their speed can also work against them. A system that makes thousands of requests in a short period may attract attention more quickly than a human attacker trying to remain unnoticed.
However, Google warns that these limitations could become less obvious as the technology improves.
The next generation of AI agents may become better at choosing their actions, avoiding detection, and coordinating complex operations. That is why security teams cannot rely only on traditional methods of monitoring and responding to attacks.
They will increasingly need AI-powered defensive systems capable of detecting suspicious activity and responding at a similar speed.
Google says it is already developing and using AI security technologies, including systems designed to identify vulnerabilities and help fix them before attackers can exploit them.