AntiFox: agentic AI pentest platform used to steal AWS secrets
Arctic Wolf details attacks run with AntiFox, an ARTEX fork hosted on Alibaba Cloud: IIS web shells, valid AWS credentials and stolen production secrets.

On 11 November 2026 Arctic Wolf Labs published research on attacks carried out with AntiFox, an open-source “agentic” penetration testing platform with a Chinese-language interface. Originally a fork of ARTEX version 0.1.1, it hands the typical steps of a test to AI agents: mapping exposed services, probing endpoints for vulnerabilities, exploitation. BleepingComputer, Security Affairs and The Hacker News covered the research; heise Security places the topic within a CERT-EU warning about a rapid rise in attacks using agent frameworks such as ARTEX.
What happened
According to the researchers, attackers run their own AntiFox instances on Alibaba Cloud and used them against internet-facing web services and the AWS environments behind them: vulnerability exploitation, web shell deployment, credential validation, configuration discovery and AWS secret harvesting. In some cases they obtained production secrets; in another, they read, copied and deleted S3 objects.
AntiFox is still available on GitHub, now at version 2.0, which improves how findings are documented and reported. Its publishers forbid use against live systems and for illegal activity but, according to BleepingComputer, describe no technical controls that enforce those restrictions.
Amazon told BleepingComputer that AWS services were not compromised and operated as designed: the attackers used credentials taken from customers’ already compromised applications to access their accounts.
Technical details
Arctic Wolf observed two entry points: an internet-facing IIS application, through a vulnerability the researchers could not identify, and a valid AWS credential obtained by unknown means. The sources cite no CVE.
IIS chain. The IIS worker process w3wp.exe created an ASPX web shell, used for:
- reconnaissance: current user, files in the application folder, permissions via
icacls, a write test in the site root; - reading
web.configwith PowerShell to find database connection strings and application settings, and checking the debug and error log folders; - an attempt to connect to an internal production database with the credentials found;
- copying configuration and log files into a web-accessible upload folder.
AWS chain. Access came from the Alibaba Cloud infrastructure using valid identities: IAM users and, in some cases, an assumed EC2 role. First came identity and permission checks (GetCallerIdentity, GetUser, ListUsers, ListRoles, ListPolicies, ListKeys) and an attempt to obtain an Amazon ECR authorisation token; then rapid enumeration with List, Get and Describe operations across IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS, ECS, EKS, Redshift, Athena, Glue, Systems Manager, API Gateway, CloudWatch, CloudFormation, DynamoDB and Bedrock. In one intrusion the same source cycled through several identities, sending hundreds of requests in about 30 minutes.
Almost every call returned AccessDenied, but the activity then focused on the resources that remained reachable: some sequences ended with a successful GetSecretValue on a production secret, others targeted S3, listing buckets, objects and versions, then reading, copying and deleting objects. This is the key point of the report: a burst of denied requests does not mean the attack failed.
MITRE ATT&CK techniques (our mapping; the sources do not publish one): T1190 (exploit public-facing application), T1505.003 (web shell), T1552.001 (credentials in files), T1078.004 (valid cloud accounts), T1087.004 and T1580 (cloud account and infrastructure discovery), T1555.006 (cloud secrets management stores), T1530 (data from cloud storage), T1485 (data destruction).
The sources publish no IP addresses, domains or hashes.
Impact for European and Italian companies
The case combines two widespread weaknesses: secondary web applications on IIS, often less closely watched, and long-term AWS keys written into configuration files. An agent-driven tool exploits them effortlessly: it tries almost twenty AWS services in half an hour and stops wherever it finds an open door. As Dark Reading puts it, criminal use of these tools is now a given, and companies need a defence plan right away.
For NIS2 entities, access to production secrets or deletion of data on S3 is an incident to assess for notification to the national CSIRT (CSIRT Italia in Italy); if the objects contain personal data, GDPR notification duties apply as well.
What to do now
- Hunt for the IIS chain:
w3wp.exelaunchingcmd.exe,powershell.exe,whoami.exe,icacls.exeor archiving tools; new.aspxfiles in upload, temporary and image-conversion folders; configuration or log files saved there with.txtor image extensions; PowerShell commands readingWebConfigurationManager,ConnectionStringsorAppSettings. - Hunt for the AWS chain in CloudTrail: activity from hosting providers the identity has never used, a single IP using many keys, users or roles in a short time, bursts of
List,GetandDescribecalls across unrelated services, especially from Boto3 or Botocore. Review GuardDuty findings for anomalous IAM user behaviour, credential access and discovery. - Correlate both sides: suspicious IIS activity followed shortly by unusual AWS access is the strongest signal. Within bursts of denials, look for the calls that succeeded.
- If you find traces: isolate the server, preserve logs and evidence, remove malicious files, revoke the AWS sessions and keys involved and rotate every credential stored in the application’s files, database credentials included.
- Limit the possible damage: no long-term access keys in configurations, IMDSv2 enforced on EC2 instances, least-privilege permissions for applications, versioning and Object Lock on important S3 buckets, alerts on access to secrets.
- Inventory and patch: list your exposed applications and prioritise those that accept uploads, convert files or process documents.
How ARXDOME helps
ARX Expose shows what an attacker sees from outside, including vulnerable services and accounts exposed in breaches, while ARX Scout verifies which ports are actually open. ARX Sentry brings server logs into the SOC pages, where you can look for suspicious executions such as a shell launched by the IIS process. ARX Horizon enriches every address with country, operator and cloud network origin, which helps spot connections from unusual hosting providers.
Last updated:


