Yaamlabs
Threat intel

AntiFox AI agents reached AWS production secrets

AntiFox, a Chinese-language fork of the ARTEX pentest tool, was linked to AWS intrusions that read a production secret in about 30 minutes. What to check.

By Yaali. November 13, 2026, 6 min read, Threat intel, Identity.

Cover illustration on a plain navy gradient, with the Yaamlabs logo and the text: AntiFox AI agents reached AWS production secrets, 1 secret stolen in about 30 minutes

Arctic Wolf Labs reported this week that self-hosted instances of AntiFox, an artificial intelligence (AI) penetration-testing platform, were linked to real intrusions that reached Amazon Web Services (AWS) production secrets. AntiFox began as a fork of the open-source ARTEX agent framework, version 0.1.1, with a Chinese-language console for launching AI agents that automate security-testing tasks. Its repository was created on October 9, 2026. The attacker infrastructure that served those AntiFox consoles ran on Alibaba Cloud.

You are exposed if an internet-facing application holds AWS credentials in a configuration file, or if any identity in your account can call AWS Secrets Manager more widely than its job needs. In one case Arctic Wolf described, an operator ran hundreds of AWS application programming interface (API) calls across several identities in roughly 30 minutes. Most were refused, but at least one request to read a production secret went through. There is no reported flaw in AWS itself. The attackers signed in with valid credentials, so this is an identity and permissions problem, not a patch.

AntiFox-linked AWS intrusion in four steps: an IIS worker process writes an ASPX web shell, the operator reads AWS credentials from web.config, several identities make hundreds of API calls in about 30 minutes, and one GetSecretValue call against a production secret succeeds

How it works

An agentic AI pentest tool takes a target and a goal, then lets a group of language-model agents plan steps, run commands and react to what comes back, with far less typing from the operator than a manual engagement. ARTEX, the framework AntiFox is built on, was published by a Chinese developer in July 2026 and passed 12,000 forks before going closed source in October. The design that makes such a tool useful to a tester, trying many things quickly and adapting on failure, also serves an intruder who runs it on their own server and aims it at someone else. AntiFox is a dual-use framework being misused, not a new malware family.

The AWS break-in did not start with an exploit against AWS. It started on a web server. Arctic Wolf saw an internet-facing application running Microsoft Internet Information Services (IIS), the Windows web server, get compromised through a vulnerability it could not identify. The IIS worker process, w3wp.exe, then wrote an ASPX web shell to the site, with file names matching the pattern pwn*.aspx. A web shell is a small script dropped into a web root that lets an attacker run commands through ordinary web requests. From there the operator read the application's web.config file and pulled out the credentials stored in it, including connection strings and AWS keys.

Those keys open the door to the cloud account. AWS Secrets Manager stores secrets such as database passwords and API tokens, and its GetSecretValue call returns the plaintext of a stored secret to any identity allowed to read it. That is why it is the prize: an attacker who can read even one secret often lands a database password or a third-party key that unlocks the next system. In this intrusion most calls came back denied, which is what a least-privilege setup should do, but at least one GetSecretValue against a production secret succeeded.

What attackers are doing

Arctic Wolf attributes the activity by infrastructure. The Alibaba Cloud hosts that drove the attacks also served AntiFox interfaces, with a console titled in Chinese as an autonomous penetration-testing console. That link ties the operations to AntiFox-linked tooling, but it does not prove an AI agent independently carried out every action, because the reporting does not separate calls issued by an agent from commands an operator ran by hand. Treat the "autonomous" framing with that caveat in mind.

The cloud phase was fast and noisy. Starting from a valid credential obtained by undetermined means, the operator cycled through several AWS identities, including at least one role assumed from an EC2 instance, and fired hundreds of API calls at more than a dozen services in about half an hour. Those services included Secrets Manager, Simple Storage Service (S3), EC2, Lambda and Amazon Bedrock. Most calls were refused with AccessDenied, the pattern you would expect from an identity probing for whatever it can reach.

One link is weaker than the rest. Arctic Wolf calls the connection between the IIS web shells and the later AWS activity a hunting hypothesis rather than a confirmed chain. It fits, because a compromised web application is a common place to find cloud credentials, but the reporting does not establish that every AWS intrusion began on an IIS server or explain how each credential was obtained. AWS told reporters its own services were not compromised and operated as designed, and that the activity used credentials taken from customers' own applications.

What to do

AWS hardening and detection checklist against AntiFox-style activity: scope GetSecretValue per role and resource, alert on GetSecretValue spikes and AccessDenied bursts across identities, find and remove IIS web shells, rotate exposed keys and move to short-lived roles, and turn on GuardDuty with non-human identity monitoring

  1. Scope secretsmanager:GetSecretValue tightly. No identity should be able to read every secret. Grant the action per role and per secret, using resource ARNs or resource tags in the policy so a role reads only the secrets its workload owns. Review which principals hold a blanket secretsmanager:* or Resource: "*" and cut them back.
  2. Alert on the two signals from this intrusion. In AWS CloudTrail, raise an alert when a single identity makes an unusual volume of GetSecretValue calls, and when one or several identities produce a burst of AccessDenied responses in a short window. Enumeration against a wall of denials, then one success, is the shape to catch.
  3. Find and remove IIS web shells. Look in your IIS web roots, under inetpub, for .aspx or .ashx files you did not deploy, especially recent ones with odd names. Turn on IIS request logging if it is off. Watch for w3wp.exe spawning cmd.exe or powershell.exe, which a normal web application almost never does, and treat any such process tree as suspect.
  4. Rotate what leaked, then stop storing long-lived keys. Rotate any AWS access key that lived in web.config, application settings or source, and revoke the active sessions tied to it, since rotating a key does not end a session that already assumed a role. Move applications to short-lived credentials from an IAM (Identity and Access Management) role attached to the instance or workload, so there is no static key in a file to steal.
  5. Turn on Amazon GuardDuty and anomaly detection. GuardDuty flags credential use from an unexpected location, instance credentials used off the instance, and anomalous API patterns, all of which appeared here. Pair it with your CloudTrail alerting rather than relying on either alone.
  6. Account for non-human identities. IAM roles, service accounts and workload identities make most of the calls in a cloud account and rarely get reviewed. Inventory them, scope each to what it needs, and alert when one reaches for secrets or services outside its normal behaviour.

The wider lesson

Hundreds of calls across several identities in half an hour is faster than a human analyst can triage a single alert, whether an agent or a script drove it. The controls that hold at that pace do not depend on reaction time: a GetSecretValue permission scoped so tightly that enumeration returns nothing useful, and credentials short-lived enough that a key pulled from a config file has already expired.

Our cloud and Kubernetes security reviews check which identities can read which secrets and where long-lived keys still sit in applications, and our security operations work builds the CloudTrail and GuardDuty detections that catch an enumeration burst while it is happening. If you are not sure which roles can call GetSecretValue today, open the chat and Yaali, our AI agent, will pass the question to an engineer.


Sources: Cloud Security Alliance Labs research note, Cryptika, GBHackers, Aviatrix Threat Research Center, Datadog Security Labs on ARTEX in AWS, ThreatAFT.

Read next

Back to the blog, or tell us about your system in the chat. Yaali, our AI agent, answers first and brings in an engineer.