Artificial intelligence agents are increasingly being developed to perform complex tasks with limited human supervision. While much of this technology is being designed for productivity, software development and business automation, cybersecurity researchers are also finding evidence that similar capabilities can be incorporated into cybercrime operations.

A recently uncovered operation linked by researchers to the French-speaking cybercrime group BlackHatSect0r && DXQRTXX provides a striking example. Researchers examining exposed infrastructure found an AI agent connected to a DeepSeek model alongside automated scanning, credential-harvesting and phishing tools.

According to the researchers, the operation was capable of scanning large numbers of internet-facing systems, identifying exposed information and feeding results into a custom platform used by the operators.

The exposed server contained approximately 4.9GB of material across 9,299 files. Researchers reported finding the DXSCAN platform, phishing tools, extortion-related material and a repository containing 16,834 credentials.

The scale of the scanning operation was also significant. Researchers said approximately 2.76 million domains had been queued, with the system reaching more than 726,000 hosts and generating approximately 1.37 million IP addresses during the campaign.

What makes the case particularly relevant to the growing discussion around AI security is how artificial intelligence was incorporated into the workflow.

The operators reportedly used a Nous Research Hermes agent connected to a DeepSeek model. According to the investigation, safeguards within the agent were disabled and several background workers were configured to continue scanning, collecting information and reporting results without requiring constant human interaction.

This represents an important development in the relationship between AI and cybersecurity. Attackers do not necessarily need artificial intelligence to discover entirely new vulnerabilities for the technology to increase their capabilities.

Instead, AI agents can potentially help automate repetitive stages of existing cyber operations.

Tasks such as reconnaissance, identifying exposed systems, processing large quantities of information and organising results have traditionally required substantial time or custom automation. Agentic systems could allow some of these activities to operate more continuously and with less direct supervision.

In the BlackHatSect0r case, researchers said the DXSCAN platform generated around 1,200 random addresses every ten seconds, checked common web ports and searched for more than 200 credential patterns.

When exposed configuration files, cloud keys or database credentials were discovered, the system could store the information and send reports to the operators.

Importantly, researchers did not find evidence that the operation depended on a previously unknown breakthrough vulnerability.

Although some tools were marketed as exploiting zero-day vulnerabilities, the investigation reportedly found that confirmed access instead came from familiar security weaknesses, including publicly accessible cloud storage, exposed environment or Git files, default signing keys and secrets embedded in browser-accessible code.

That distinction matters because it demonstrates how AI could amplify problems that cybersecurity teams already understand.

A forgotten configuration file or incorrectly configured cloud storage bucket may not represent a new class of vulnerability. But automated systems capable of continuously searching large numbers of targets can potentially discover those mistakes faster and at greater scale.

The credential repository uncovered during the investigation provides an indication of that scale. Researchers said it grew from 16,415 records on August 11 to 16,834 by August 18, containing database credentials, API keys, cloud access keys, GitHub tokens, email credentials and other secrets.

Some of the infrastructure was also reportedly connected to phishing activity.

One technique involved encouraging potential victims to call an attacker-controlled telephone number rather than directing them to click a malicious link. This type of social-engineering approach can present a different challenge for traditional email security tools because there may be no malicious attachment or destination link to analyse.

The broader concern is that AI agents could help cybercriminals combine several stages of an operation into increasingly automated workflows.

An agent could potentially assist with processing reconnaissance results, prioritising targets and organising information, while conventional scripts and scanning tools perform the underlying technical actions. Human operators can then remain involved where judgement, social engineering or strategic decisions are required.

This does not mean fully autonomous AI hackers have replaced human attackers.

The BlackHatSect0r case instead illustrates a more immediate cybersecurity challenge: AI can become another layer of automation inside an already established cybercrime operation.

That distinction is important when evaluating emerging AI threats. Cybercriminals do not need a completely autonomous system capable of independently compromising organisations from beginning to end. Even partial automation could increase the speed, persistence and scale of existing attack techniques.

For defenders, many of the countermeasures remain familiar.

Organisations should ensure cloud storage is not unintentionally publicly accessible and prevent sensitive configuration files, development repositories and debugging interfaces from being exposed to the internet. Credentials and API keys should also be regularly reviewed and rotated when exposure is suspected.

Secrets used for authentication or token generation should remain on protected server-side infrastructure rather than appearing inside browser-accessible code. Default credentials and easily guessed signing keys should also be eliminated.

Continuous monitoring becomes increasingly important when attackers themselves are becoming more automated. Security teams need the ability to identify unusual scanning behaviour, unexpected authentication activity and attempts to access configuration files before exposed information can be used for further attacks.

The case also demonstrates why organisations should focus on basic security hygiene alongside emerging AI threats.

Artificial intelligence may make cyber operations faster, but attackers can still succeed because of familiar weaknesses such as exposed credentials, poor cloud configuration and insufficient access controls.

As AI agents become more capable, the cybersecurity challenge will therefore involve more than defending against entirely new forms of attack. Organisations will also need to prepare for existing techniques being executed with greater speed and automation.

The emergence of AI-assisted cybercrime infrastructure suggests that the same technologies transforming legitimate business processes can also change how malicious operations are organised.

For cybersecurity teams, the lesson is not that every future attack will be controlled by an autonomous AI. It is that attackers now have another technology capable of helping them scale repetitive work.

As agentic AI continues to develop, reducing easily exploitable weaknesses and improving automated defensive monitoring could become even more important. When attackers can search for mistakes at machine speed, organisations have less room to leave those mistakes exposed.

The technical details and figures above are based on Cyber Security News' reporting of research into the exposed BlackHatSect0r && DXQRTXX infrastructure. The attribution and claims concerning the group's activities should therefore be understood as findings reported by the researchers rather than independently verified conclusions.