
BleepingComputer reported that attackers are hosting AntiFox, an open-source agentic pentest platform, on Alibaba Cloud and using it against internet-facing web services and the AWS accounts behind them. Arctic Wolf is the source for these AntiFox AWS attacks: webshells, credential checks, and AWS secret harvesting on the hosts it tracked.
Key takeaways
- Arctic Wolf tied self-hosted AntiFox on Alibaba Cloud to web exploitation, webshells, and AWS secret harvesting, with no CVE or named victims.
- One case: an ASPX webshell on internet-facing IIS, then hundreds of authenticated AWS API calls across multiple identities in about 30 minutes. Most were denied. Some returned production secrets.
- Amazon said AWS itself was not compromised; the credentials came from customers' applications.
- AntiFox 2.0, a fork of ARTEX 0.1.1, is still on GitHub. These incidents are not attributed to Vapor Panda.
What did Arctic Wolf tie to the AntiFox AWS attacks?
Arctic Wolf associated Alibaba Cloud systems hosting AntiFox interfaces with confirmed attacks on internet-facing web services and the AWS environments attached to them.
AntiFox is an open-source, self-hostable, Chinese-language agentic penetration-testing platform. It began as a fork of ARTEX 0.1.1 and remains on GitHub at version 2.0. Users deploy AI agents to automate routine pentest tasks such as mapping exposed services and probing endpoints.
The same infrastructure supported vulnerability exploitation, webshell deployment, credential validation, configuration discovery, and AWS secret harvesting. Operators made authenticated AWS API calls with valid identities, including IAM users and, in some cases, assumed EC2 roles.
The reporting names no CVE, no victim, and no count of affected accounts. Arctic Wolf describes several recently observed intrusions rather than one named breach.
How did the attack move from an IIS webshell to AWS?
In the intrusion Arctic Wolf highlighted, operators deployed an ASPX webshell on an internet-exposed Microsoft IIS server. From the shell they ran OS commands, read application settings and database connection strings, and copied configuration and log files into a web-accessible directory for retrieval over HTTP. Webshells often run as the application pool identity and can read web.config without further escalation; the source does not say which account this shell used.
They then cycled through multiple AWS identities and issued hundreds of API requests in about 30 minutes. Most calls were blocked. Some succeeded, and production secrets were obtained. In a separate incident a valid AWS identity read, copied, and deleted S3 objects.
A web application penetration testing pass that stops at the login form misses this chain. What matters is which cloud identity is usable once those settings and connection strings are exposed.
Why is this not an AWS platform breach?
Amazon told BleepingComputer that AWS services were not compromised. The activity used credentials taken from customers’ compromised applications to reach their AWS accounts, and customers were pointed to credential-remediation guidance or AWS Support.
- Least privilege on every identity the application can use. AI agent least privilege is the design choice behind that control.
- Require IMDSv2 on EC2. Version 1 answers a plain request from any process on the instance, including a webshell. Version 2 demands a session token first.
- Keep long-term access keys out of application configuration. Those keys do not expire by themselves, so workloads should use roles.
- Turn on Amazon GuardDuty and watch for InstanceCredentialExfiltration, the finding for EC2 role credentials used off the instance.
The playbook stays credential-centric: revoke affected sessions and keys, and rotate every secret exposed in application files, including database passwords.
AntiFox, ARTEX, and Vapor Panda: what is confirmed?
AntiFox is a fork of the open ARTEX 0.1.1, not the later closed-source ARTEX AI product, which went closed-source after attacks on South Korean banks. CrowdStrike linked ARTEX AI to a campaign in which the Chinese group Vapor Panda targeted more than 200 organizations in 14 countries. This reporting does not attribute the AntiFox incidents to that group; the operators are unnamed.
AntiFox 2.0’s publishers prohibit live systems, real penetration tests, and destructive or illegal activity, with no technical control that enforces the ban. A written ban does not bind an operator who already self-hosts the agents. AI agent penetration testing checks tool use, goal hijack, and credential scope for this class of automation.
What should a security lead hunt and scope?
Correlate IIS webshell activity with unusual AWS access, then review the successful API calls inside a burst of denials. CloudTrail records authenticated management events, denials included, with identity ARN, source IP, and user agent when the trail logs them. From the webshell timestamp, group AccessDenied events by sourceIPAddress and userIdentity.arn, then list event names in that window with no error. The highlighted intrusion was hundreds of calls in about 30 minutes: that is the burst to measure against. S3 reads, copies, and deletes are data events and appear only if that logging was on.
If the source address is an Alibaba Cloud network and an EC2 role is used off the instance, look for GuardDuty InstanceCredentialExfiltration and any identity calling from anywhere but the instance.
AWS penetration testing for this pattern covers identity cycling, role use after a web foothold, IMDSv1 versus IMDSv2, keys left in config, and what a partly allowed API burst can still take. Brief leadership on what is confirmed: a web compromise, valid cloud identities, partial API success, and production secrets taken.
FAQ
What is AntiFox and who is using it?
AntiFox is an open-source, Chinese-language agentic pentest platform, forked from ARTEX 0.1.1 and published on GitHub at version 2.0. Unidentified operators hosted their own instances on Alibaba Cloud and used them against internet-facing web services and related AWS accounts.
Were AWS services compromised in the AntiFox attacks?
No. Amazon said AWS services were not compromised. The activity used credentials taken from customers' compromised applications.
Is AntiFox linked to Vapor Panda?
AntiFox forked the open ARTEX 0.1.1. CrowdStrike linked the later ARTEX AI to a campaign in which Vapor Panda targeted more than 200 organizations in 14 countries. This reporting does not attribute the AntiFox incidents to that group.
What is the first remediation step after exposed application secrets?
Revoke the affected AWS sessions and keys, and rotate every credential that sat in the application files, including database passwords. Then review successful CloudTrail events inside the burst of denials.
If your web tier can become an IAM user or an EC2 role, that chain belongs in the test scope. agent::offense runs AWS penetration testing against the identities an application compromise can actually reach.