Artificial intelligence tools designed for automated cybersecurity testing are increasingly being used by malicious actors to carry out real-world attacks. One such platform, AntiFox, has been linked to intrusions into corporate web systems, credential theft, and access to confidential information in Amazon Web Services (AWS) cloud infrastructure. The Hacker News reported this, citing research by Arctic Wolf Labs.

According to the researchers, in several identified incidents, Alibaba Cloud servers hosting AntiFox interfaces were used to attack internet-facing web services and their associated AWS environments. The attackers combined the exploitation of web application vulnerabilities with the installation of web shells, which allowed them to execute commands on compromised servers, examine system configurations, and extract credentials.

AntiFox is an autonomous penetration testing platform available on GitHub and built on the ARTEX project. It uses a system of AI agents that can independently carry out the main stages of testing, from reconnaissance and vulnerability discovery to exploitation and report generation. Its developers position the tool as a solution that can not only identify security issues but also document the results so that engineers can reproduce the vulnerabilities discovered.

How attackers gained access to data

In one scenario described by Arctic Wolf, the initial intrusion occurred through a vulnerable internet-facing application on a Microsoft IIS server. The attackers then created an ASPX web shell using the w3wp.exe process, which runs IIS web applications.

After gaining the ability to execute commands on the server, the attackers searched for configuration files and their backups, examined application settings and database connection strings, checked error logs, and attempted to connect to an internal production database using the credentials they found. Configuration files and logs were also copied to an internet-accessible directory.

The researchers note that the exact vulnerability that enabled code execution on the IIS server has not yet been identified. In other cases, the attacks began with valid AWS credentials whose origin could not be determined either.

Using these credentials, the attackers checked access permissions and explored dozens of Amazon services, including IAM, EC2, S3, Lambda, Secrets Manager, KMS, and RDS. Most requests were denied access, but some operations succeeded. In at least several cases, the attackers were able to retrieve the contents of production secrets through AWS Secrets Manager. In other incidents, they used valid accounts to search for objects and perform operations in S3 storage.

Commenting on The Hacker News report, an AWS spokesperson said that the company's services had not been compromised and were operating normally. According to the spokesperson, the incidents described involved attackers using customer credentials stolen from compromised applications. AWS recommended that customers restrict access permissions, avoid storing long-term keys in application configurations, and use security monitoring tools, particularly Amazon GuardDuty.

AI agents increase the speed and scale of cyberattacks

The AntiFox incidents fit into a broader trend of autonomous AI tools being used to conduct cyberattacks. These systems can independently perform sequences of operations that previously required a specialist's continuous involvement: searching for vulnerable servers, testing discovered credentials, exploring infrastructure, and moving from one compromised resource to another.

The Hacker News also cites other examples. In particular, Check Point researchers discovered more than 2,500 systems hosting ARTEX. In ten cases, misconfigurations left entire databases accessible. According to the researchers' observations, operators of such systems often need only specify an end goal; the agents perform the subsequent actions independently, while human involvement is limited to checking progress and adjusting tasks.

In individual campaigns, ARTEX tools were used to steal cryptocurrency and Apple code-signing certificates, hack gaming accounts, and conduct DDoS attacks. During one operation, attackers obtained more than 300 Apple certificates and their associated passwords. Three certificates passed full validation and could potentially have been used to sign arbitrary iOS applications.

Experts assess that the key change AI agents bring to cyberthreats primarily concerns the scale and speed of attacks. Even attackers with limited technical skills can task agents with investigating numerous targets in parallel and exploiting known vulnerabilities, reducing the need to manage each stage manually.

As a result, tools originally intended for legitimate security testing are becoming a readily available means of automating attacks. Their ability to independently build long chains of actions increases the risk that even small groups of attackers will be able to conduct complex operations with less time and fewer resources.