← All articles

AI pentesting tool tied to AWS production secret access

AI pentesting tool tied to AWS production secret access

AI pentesting tools were linked to attacks that stole Amazon Web Services (AWS) credentials and reached production secrets, according to Arctic Wolf Labs. The activity involved AntiFox, an emerging agentic penetration testing platform, and included web exploitation, web shell deployment, cloud service discovery and access to storage.

Arctic Wolf said Alibaba Cloud systems hosting AntiFox interfaces were associated with confirmed malicious activity against internet-facing web services and related AWS environments. A web shell is a hidden program placed on a web server that lets an intruder run commands remotely. In the observed activity, attackers used these tools to discover configurations, validate credentials, enumerate cloud services, access secrets and perform storage operations.

AntiFox is available on GitHub and is described as an artificial intelligence driven autonomous penetration testing platform built on ARTEX. At the time described, the project had been forked 251 times and starred 474 times. Its documentation describes a multi-agent system in which planner, worker and main-agent components move through reconnaissance, attack-surface discovery, exploitation, evidence collection and report delivery using a global asset graph and a task-specific exploration graph. The platform is designed to continue beyond discovery and turn findings into reproducible reports.

In some incidents, Arctic Wolf traced initial access to a compromised internet-facing Internet Information Services (IIS) application. IIS is Microsoft software that hosts websites and web applications. Attackers used the IIS worker process, w3wp.exe, to create an ASPX web shell and execute operating system commands. They searched for web.config files and configuration backups, read application settings and database connection strings, inspected debug and error-log directories, and tried to connect to an internal production database using a credential taken from the application configuration. They also copied configuration and log files into a web-accessible upload directory.

Commands run through w3wp.exe used whoami to identify the account running the process, listed the application directory, checked permissions with icacls and tested whether files could be written into the web root. Arctic Wolf said that creating the ASPX web shell through w3wp.exe and then using the same process for command execution suggested that the affected web application was the route to compromise. The precise vulnerability that enabled code execution on the IIS server remains unknown.

Other attacks began with a valid AWS credential obtained by means that have not been determined. The attackers abused valid identity and access management (IAM) identities, which control permissions for AWS users and services, to validate credentials and query dozens of services. Those included IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS, ECS, EKS, Redshift, Athena, Glue, Systems Manager, API Gateway, CloudWatch, CloudFormation, DynamoDB and Amazon Bedrock.

Most of the authenticated requests returned AccessDenied, meaning the identity did not have permission to complete the requested action. However, some discovery and secret-access operations succeeded. In at least some cases, attackers moved from listing Secrets Manager resources to successfully calling GetSecretValue against a production secret. Elsewhere, a valid AWS identity was used to discover S3 resources and perform object operations through application programming interface (API) calls.

Arctic Wolf said agentic frameworks make penetration-testing methods more accessible to threat actors and increase the speed and consistency of their use in live attacks. AWS said its services were not compromised and operated as designed. The company said the incidents involved credentials taken from customers' compromised applications, and recommended least-privilege permissions, requiring IMDSv2 on EC2 instances, keeping long-term access keys out of application configurations and enabling Amazon GuardDuty. Customers that suspect unauthorised activity were directed to AWS guidance for remediating potentially compromised credentials or to AWS Support.

The disclosure forms part of a wider pattern involving AI-powered penetration testing tools such as ARTEX. Researchers said these systems can automate multiple stages of an intrusion with limited human intervention. One example cited by VECERT, called Operation Ghost-Campus, used AI agents to maintain access to large language models and was attributed to a Chinese threat actor. VECERT said the operation created accounts on fofa.info, scanned for misconfigured servers, extracted OpenAI and DeepSeek API keys, checked their financial balances and placed them into an attacker development environment. The stolen computing resources supported an interactive GoTTY terminal called Pi, an AI Coding Assistant, as well as bots used for academic fraud and for extracting financial and personal data belonging to more than 5,000 students at the Fengxian Campus.

Censys counted 1,465 hosts running an exposed ARTEX console on November 4, 2026, with most located in the United States, Hong Kong, China, Japan, Singapore and South Korea. Censys noted that not every exposed instance was necessarily malicious. Separately, Check Point Research identified 2,557 systems hosting ARTEX, including 10 misconfigured systems that exposed their full databases. Check Point said operators often supplied an objective and left the agent fleet to work, with human involvement mainly limited to checking progress, asking whether a target had been compromised or changing priorities.

The attacks observed by Check Point were mainly intended to steal cryptocurrency, obtain Apple code-signing certificates, compromise gaming accounts and launch denial-of-service attacks against gambling websites. Reported weaknesses included weak or hard-coded passwords, missing authentication, insecure direct object references and remote code execution. In the campaign involving Apple certificates, ARTEX agents searched compromised systems for certificate databases, then checked passwords, private keys, Apple revocation status, expiry dates and provisioning profiles. More than 300 certificates and their passwords were estimated to have been harvested, although only three passed the full validation process and could sign arbitrary iOS applications.

The wider ecosystem includes about 45 open-source agentic attack frameworks that connect large language models to tools and act on their output.

Also available in: DeutschItalianoShqip