High 5 Hunt LeadsIOCsBleepingComputer

AntiFox AI Pentest Agents Turn a Hacked Web App Into AWS Secrets

Arctic Wolf says self-hosted copies of AntiFox, an open-source AI pentesting platform, were tied to intrusions that went from IIS web shells to stolen AWS production secrets and deleted S3 data.

Another Chinese-language AI pentesting kit has found its way into real break-ins. BleepingComputer reports that Arctic Wolf Labs tied self-hosted copies of AntiFox, an open-source agentic penetration-testing platform forked from ARTEX, running on Alibaba Cloud, to several intrusions against internet-facing web servers and the AWS accounts behind them. In some cases the attackers walked away with production secrets; in another they read, copied and deleted S3 data. Arctic Wolf names no threat group.

For defenders

The web side: on a compromised IIS server, w3wp.exe wrote an ASPX web shell and ran whoami, icacls and directory listings (T1505.003 (Server Software Component: Web Shell)). Arctic Wolf couldn’t identify the bug used to get code execution (T1190 (Exploit Public-Facing Application)). PowerShell then parsed web.config connection strings and app settings with .NET configuration classes (T1552.001 (Unsecured Credentials: Credentials In Files)), tried the database with a pulled credential, and copied config and log files into a web-reachable upload folder as cfg*.txt, wc*.txt, el*.txt.

The cloud side: valid IAM users and, sometimes, an assumed EC2 role (T1078.004 (Valid Accounts: Cloud Accounts)). One source cycled through several identities and fired hundreds of List/Get/Describe calls across IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS, EKS, Bedrock and more in about 30 minutes (T1526 (Cloud Service Discovery)). Most were denied; buried in the burst was a successful GetSecretValue on a production secret (T1555.006 (Credentials from Password Stores: Cloud Secrets Management Stores)). Elsewhere: GetObject, CopyObject and DeleteObject on S3 (T1530 (Data from Cloud Storage)). User agent Boto3/Botocore 1.43.125.

Context: Amazon says AWS itself wasn’t compromised; the credentials came from customers’ hacked applications. AntiFox is still on GitHub, now at 2.0. It follows the ARTEX cases we’ve covered, from the Korean bank breaches to VAPOR PANDA. Arctic Wolf warns the cloud addresses may be reassigned.

Network
47.79.4[.]33 (AntiFox-linked source, Alibaba Cloud)
8.208.101[.]210 (AntiFox-linked source, Alibaba Cloud)
AntiFox console on TCP/8787 · banner hash c370d03542ed2ec696f7337940b14652e98e9bab

Host
w3wp.exe → cmd.exe / powershell.exe / where.exe · cmd.exe → whoami.exe / icacls.exe
upload directory: pwn*.aspx · cfg*.txt · wc*.txt · el*.txt

What it means for you

If you don’t run a website on Windows servers or use AWS, there’s nothing for you to do here. If you do, this is the pattern to fear: one hacked web app gives up the passwords and cloud keys stored in its settings file, and an AI tool then tries all of them against your cloud account faster than a person could. Most of its attempts fail, which is exactly why a quick look at your logs can miss the ones that worked.

Do this now

  1. Keep AWS keys out of web.config and app settings; use instance roles with IMDSv2 and least-privilege policies.
  2. Alert when your IIS worker process (w3wp.exe) starts cmd.exe or PowerShell, or writes .aspx files.
  3. Review CloudTrail for bursts of denied calls from new hosting networks and check which calls in that burst succeeded.
  4. If you find a web shell, rotate every credential in that app’s config files and revoke the AWS sessions and keys they touched.

Source article: Hackers use AntiFox AI agents in AWS credential-theft attacks
Publisher: BleepingComputer
Published: Nov. 12, 2026
URL: https://www.bleepingcomputer.com/news/security/hackers-use-antifox-ai-agents-in-aws-credential-theft-attacks/

Executive Summary

Arctic Wolf Labs is tracking malicious use of self-hosted AntiFox, an open-source Chinese-language agentic penetration-testing platform originally forked from ARTEX 0.1.1, from Alibaba Cloud infrastructure; no threat actor is named, and BleepingComputer places it alongside earlier criminal and espionage use of ARTEX, including CrowdStrike’s VAPOR PANDA reporting. In several intrusions, AntiFox-linked hosts were tied to confirmed malicious actions against internet-facing web services and the AWS environments behind them. On one compromised IIS application, the w3wp.exe worker process created an ASPX web shell (the exploited vulnerability was not identified), ran whoami, icacls and directory discovery, used PowerShell and .NET configuration classes to parse web.config connection strings and application settings, attempted an internal production database login with an extracted credential, and staged configuration and log files in a web-accessible upload directory. Separately, valid IAM users and an assumed EC2 role, some obtained by unknown means, were used from the same infrastructure to validate credentials and enumerate dozens of AWS services in bursts of hundreds of calls within half an hour; most calls were denied, but successful operations inside the bursts led to a GetSecretValue on a production secret, and in another case to S3 GetObject, CopyObject and DeleteObject activity. The objective is credential and secret theft leading to cloud data access, with destructive object deletion observed. Hunt leads, false positive expectations and indicators are detailed below; Arctic Wolf published hunting guidance and indicators but no detection rules.

ATT&CK chain: Initial Access → Persistence → Discovery → Credential Access → Collection → Impact

Hunt Lead Grid

Hunt Lead Level Hunt Lead Name ATT&CK Risk of False Positives
1 (Junior) AntiFox Source Infrastructure and SDK User Agent T1078.004 (Valid Accounts: Cloud Accounts) Medium. The two addresses are on shared public cloud and may be reassigned, and Boto3 is common; the lead anchors on the exact addresses in CloudTrail and IIS logs, and uses the user agent only together with a novel hosting source.
1 (Junior) ASPX Web Shells and Staged Config Files in Upload Directories T1505.003 (Server Software Component: Web Shell) Low. Upload, temp and image-conversion folders should not contain .aspx files or .txt copies of configuration and logs; deployments write to application folders through a release process, not via w3wp.exe.
2 (Pro) IIS Worker Spawning Shells and Discovery Tools T1505.003 (Server Software Component: Web Shell), T1190 (Exploit Public-Facing Application) Medium. A few applications legitimately shell out from w3wp.exe; the lead allowlists those by application pool and exact command line and flags whoami, icacls and where.exe, which applications almost never run.
2 (Pro) PowerShell Reading web.config Through .NET Configuration Classes T1552.001 (Unsecured Credentials: Credentials In Files) Low. Administrators rarely script WebConfigurationManager or ConnectionStrings reads on production web servers, and almost never from a w3wp.exe process tree.
3 (Elite) Denied API Burst Hiding a Successful GetSecretValue T1526 (Cloud Service Discovery), T1555.006 (Credentials from Password Stores: Cloud Secrets Management Stores), T1530 (Data from Cloud Storage) Medium. Security scanners and CSPM tools also enumerate many services quickly; the lead requires a novel source network, multiple identities from one IP, and a successful secret or object read inside a burst of AccessDenied errors.

AntiFox Source Infrastructure and SDK User Agent

Arctic Wolf published two Alibaba Cloud addresses linked to AntiFox activity, 47.79.4[.]33 and 8.208.101[.]210, together with the AntiFox web console fingerprint (title “AntiFox — 自主渗透测试控制台”, banner hash c370d03542ed2ec696f7337940b14652e98e9bab, commonly on TCP/8787) and the AWS SDK user agents Boto3/1.43.125 and Botocore/1.43.125 seen on the malicious API calls. The detection logic is a retrospective search for either address as the sourceIPAddress in CloudTrail or as the client in web server, WAF and firewall logs, and a second search for AWS API calls carrying that Boto3/Botocore version from any autonomous system the calling identity has never used before.

Use CloudTrail management and S3 data events across all regions, IIS and load-balancer access logs, WAF logs and perimeter firewall logs. Look back at least 90 days, since the credentials used may have been stolen well before the API activity. Scope to internet-facing web servers and identities with programmatic access. Arctic Wolf notes these addresses are public-cloud infrastructure that may be reassigned or shared, so a hit from months ago needs the behavioral leads to confirm it, and blocking all Alibaba Cloud ranges is the wrong control for most organizations. The console fingerprint is useful for internet-scanning platforms to find other AntiFox deployments, not for internal logs. Any CloudTrail call from either address under one of your identities should be escalated, with the identity’s keys revoked.

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Initial Access, Persistence → T1078.004 (Valid Accounts: Cloud Accounts)

Sources: “Hackers use AntiFox AI agents in AWS credential-theft attacks”, BleepingComputer, Nov. 12, 2026, https://www.bleepingcomputer.com/news/security/hackers-use-antifox-ai-agents-in-aws-credential-theft-attacks/ (source article); “2026-11-10: Antifox Agentic Pentesting Platform Used to Steal AWS Credentials and Move Laterally in Compromised Environments”, Arctic Wolf Labs, Nov. 10, 2026, https://github.com/rtkwlf/wolf-tools/tree/main/pack_alerts/202611-antifox-alibaba-hosted-platforms-facilitate-web-cloud-intrusions (additional source, the research the source article reports)

ASPX Web Shells and Staged Config Files in Upload Directories

On the compromised IIS server, the worker process created an ASPX web shell in a web-accessible upload directory (Arctic Wolf lists pwn*.aspx and generic *.aspx there) and later copied configuration and log files into the same reachable location as cfg*.txt, wc*.txt and el*.txt so they could be downloaded over HTTP. The detection logic is a file sweep and a file-creation alert: any .aspx, .ashx or .asmx file in upload, temp, image-conversion or other user-writable web directories, any .txt or image-extension file in those directories whose contents look like XML configuration, connection strings or application logs, and any of these created by w3wp.exe.

Use EDR file-creation telemetry or Sysmon Event ID 11, file integrity monitoring on web roots, and IIS access logs for GET requests to the new files. Sweep current web roots now regardless of date and look back 90 days in telemetry. Scope to internet-facing IIS servers, prioritizing upload, document-processing and file-conversion applications, which Arctic Wolf singles out. False positives are rare in upload folders; if an application legitimately writes .aspx there, that is itself a design flaw worth fixing. Deleting the shell is the wrong first step: preserve the web root, IIS logs and memory, then remove all unauthorized files, since attackers often leave alternate shells. Any hit should be escalated and every secret in that application’s configuration treated as stolen.

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Persistence → T1505.003 (Server Software Component: Web Shell)

Sources: Arctic Wolf Labs pack alert (above), additional source; BleepingComputer (above), source article

IIS Worker Spawning Shells and Discovery Tools

After the shell was in place, commands ran through the same IIS worker process: w3wp.exe started cmd.exe, powershell.exe and where.exe, and cmd.exe ran whoami.exe to identify the execution context and icacls.exe to check permissions, followed by directory enumeration and tests of whether files could be written into the web root. The detection logic is a parent-child rule: w3wp.exe as the parent of cmd.exe, powershell.exe, pwsh.exe, where.exe or archive utilities, and any of whoami.exe, icacls.exe, net.exe or similar discovery tools appearing anywhere in a w3wp.exe process tree.

Use EDR process telemetry or Sysmon Event ID 1 with parent image and command line, and Windows Security Event 4688 with command-line auditing where EDR is absent. Baseline 30 days per application pool to find the few that legitimately shell out (some PDF converters and legacy apps do) and allowlist them by application pool identity and exact command line, never by child process name alone. Look back 90 days. Scope to every IIS server, internet-facing first. Arctic Wolf recommends isolating the server immediately when IIS spawns unexpected shells; that is the right triage, because by that point the attacker has code execution and is reading configuration. The rule generalizes to any web shell family, which makes it worth keeping long after this campaign.

Hunt Lead Level: 2 (Pro)

ATT&CK: Persistence → T1505.003 (Server Software Component: Web Shell); Initial Access → T1190 (Exploit Public-Facing Application)

Sources: Arctic Wolf Labs pack alert (above), additional source

PowerShell Reading web.config Through .NET Configuration Classes

Rather than just printing web.config, the actor used PowerShell with .NET configuration classes to parse connection strings and application settings, then enumerated debug and error-log directories and tried to log in to an internal production database with a credential it had extracted. Arctic Wolf’s hunting guidance names the strings to look for: System.Web.Configuration.WebConfigurationManager, WebConfigurationFileMap, VirtualDirectoryMapping, ConnectionStrings.ConnectionStrings, AppSettings.Settings and [IO.File]::ReadAllText. The detection logic is a PowerShell content rule on those strings, raised in severity when the PowerShell process descends from w3wp.exe or is followed by a database connection from the web server using a different account than the application pool normally uses.

Use PowerShell Script Block Logging (Event ID 4104) and Module Logging, AMSI-integrated EDR content, and database server login audit logs for the second half of the correlation. Look back 90 days. Scope to IIS and application servers. Administrators seldom script these classes on production hosts, so false positives should be few and can be allowlisted by signed script path and account. Obscuring web.config with encryption does not stop an attacker running inside the application’s own process, so the real fix is keeping cloud keys and database passwords out of the file and rotating them after any hit. Escalate any match from a w3wp.exe tree as confirmed credential access.

Hunt Lead Level: 2 (Pro)

ATT&CK: Credential Access → T1552.001 (Unsecured Credentials: Credentials In Files)

Sources: Arctic Wolf Labs pack alert (above), additional source

Denied API Burst Hiding a Successful GetSecretValue

From AntiFox infrastructure, one source cycled through several AWS identities, including IAM users and an assumed EC2 role, and generated hundreds of cross-service API requests in about 30 minutes: GetCallerIdentity, GetUser, ListUsers, ListRoles, ListPolicies, ListKeys, an attempted ECR GetAuthorizationToken, and List, Get and Describe calls across IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS, ECS, EKS, Redshift, Athena, Glue, Systems Manager, API Gateway, CloudWatch, CloudFormation, DynamoDB and Bedrock. Most returned AccessDenied, and the actor moved on to whatever stayed accessible; in some cases the sequence ended with ListSecrets followed by a successful GetSecretValue on a production secret, and elsewhere with ListBuckets, ListObjects, HeadObject and GetObject followed by CopyObject and DeleteObject. The detection logic is a windowed correlation per source IP: more than one access key or role session from a single IP within an hour, a high count of distinct services (Arctic Wolf saw dozens) with a high AccessDenied ratio, and any successful GetSecretValue, GetObject, CopyObject or DeleteObject inside that window.

Use CloudTrail management events plus S3 data events for sensitive buckets (which must be enabled in advance for object-level calls to appear), GuardDuty findings for anomalous IAM behavior and EC2 credential use from outside AWS, and ASN enrichment of sourceIPAddress. Baseline 30 days of source networks per identity so a first-seen hosting ASN stands out. Look back 90 days. Scope to identities with programmatic keys and to EC2 instance roles, since a role session used from an external IP means instance credentials were stolen. Exclude CSPM and scanner roles by role ARN, not by user agent. Do not dismiss a burst because most calls failed: Arctic Wolf’s point is that the successes are buried inside the failures. Any successful secret or object read in a matching window should be escalated; revoke sessions, rotate the secret that was read, and check S3 versioning to recover deleted objects.

Hunt Lead Level: 3 (Elite)

ATT&CK: Discovery → T1526 (Cloud Service Discovery); Credential Access → T1555.006 (Credentials from Password Stores: Cloud Secrets Management Stores); Collection → T1530 (Data from Cloud Storage)

Sources: Arctic Wolf Labs pack alert (above), additional source; BleepingComputer (above), source article

Hash Appendix

No file hashes were published in the source article or Arctic Wolf’s pack alert. The banner hash c370d03542ed2ec696f7337940b14652e98e9bab is an AntiFox web console fingerprint for internet-scanning platforms, not a file hash, and should not be loaded into file-hash blocklists.

References Appendix

  • “Hackers use AntiFox AI agents in AWS credential-theft attacks”, BleepingComputer, Nov. 12, 2026, https://www.bleepingcomputer.com/news/security/hackers-use-antifox-ai-agents-in-aws-credential-theft-attacks/ (source article, including Amazon’s statement)
  • “2026-11-10: Antifox Agentic Pentesting Platform Used to Steal AWS Credentials and Move Laterally in Compromised Environments”, Arctic Wolf Labs, Nov. 10, 2026, https://github.com/rtkwlf/wolf-tools/tree/main/pack_alerts/202611-antifox-alibaba-hosted-platforms-facilitate-web-cloud-intrusions (observed behavior, hunting guidance, indicators)
  • AntiFox project repository, https://github.com/AntiFox-AI/antifox (tool background)
  • MITRE ATT&CK Enterprise technique pages linked above, MITRE, https://attack.mitre.org/

More from the Roundup