Hunting the hunters: ARTEX, AntiFox and PentAGI
文章目录
Earlier today i asked if you want my take on ARTEX, AntiFox and PentAGI. Here it is, but as something you can actually use: what to look for and which logs to pull.
TL;DR: you hunt these through auth events, HTTP requests, endpoint processes, cloud API calls and data access. Infra, models, headers and tools all change, so don't hang everything on one IOC. Correlate the actions across those logs.
1. Start with known infrastructure IOCs
Historical IP addresses provide starting points for retrospective investigation.
| Framework / campaign | Reported IP | Where to hunt |
|---|---|---|
| ARTEX / Storm-0940 | 144.172.64[.]60 | Entra sign-ins, Microsoft Graph, WAF |
| ARTEX / Storm-0940 | 144.172.64[.]99 | Entra sign-ins, Microsoft Graph, WAF |
| ARTEX / Storm-0940 | 144.172.64[.]128 | Entra sign-ins, Microsoft Graph, WAF |
| AntiFox-linked activity | 47.79.4[.]33 | WAF, IIS, CloudTrail, firewall |
| AntiFox-linked activity | 8.208.101[.]210 | WAF, IIS, CloudTrail, firewall |
Microsoft published the ARTEX campaign IPs in October 2026. Arctic Wolf reported the AntiFox-linked infrastructure in November 2026.
2. Collect HTTP, WAF and reverse-proxy evidence
Look in Cloudflare Security Events and HTTP request logs, AWS WAF, CloudFront, ALB logs, NGINX, Apache, IIS, API gateways and Zeek HTTP logs.
| Field | Use |
|---|---|
| Source IP, ASN and country | Infrastructure and origin pivots |
| Timestamp and request ID | Sequence reconstruction |
| Hostname and destination | Target identification |
| HTTP method and full path | Enumeration and sensitive-endpoint access |
| User-Agent | Client clues and tool changes |
| Response status | Failed versus successful requests |
| Request/response size | Unusual transfers and response patterns |
| WAF rule, action and attack score | Security-control context |
| Authenticated user/session ID | Connects requests to an identity |
| URI query parameters | Parameter probing, with secrets redacted |
Look for one source or correlated identity performing:
-
Broad endpoint enumeration across many routes.
-
Repeated 401, 403, 404 or 500 responses followed by successful requests.
-
Rapid variation of HTTP methods, parameters or payload structures.
-
Access to configuration files, admin interfaces or exposed secrets.
-
A transition from reconnaissance to sensitive API operations.
Request rate alone is weak evidence because legitimate crawlers and monitoring systems can behave similarly. Enumeration followed by successful authentication or an unexpected state-changing request to the same application provides a stronger signal.
3. Collect endpoint and web-server telemetry
Look in Microsoft Defender for Endpoint, CrowdStrike, SentinelOne, Sysmon, Windows Security logs, Linux auditd, osquery and application-server logs.
Prioritize process execution, process ancestry, file creation, network connections and persistence events.
AntiFox-linked investigations included suspicious IIS files and web-server execution. Incident-derived file patterns include:
pwn*.aspx
cfg*.txt
wc*.txt
el*.txt
These patterns are not unique to AntiFox. Also look for these process relationships:
w3wp.exe -> cmd.exe
w3wp.exe -> powershell.exe
w3wp.exe -> whoami.exe
w3wp.exe -> icacls.exe
Correlate:
-
A web-server process creating or executing a script in a writable web directory.
-
An HTTP request for the same file.
-
An unusual child process launched by the web-server worker.
-
Reads, copies or staging of configuration files containing credentials.
-
Unexpected outbound traffic from the host.
Collect the file's SHA-256, creation time, full path, initiating process, user, destination address and related HTTP request IDs.
A web-server child shell combined with suspicious file activity provides stronger evidence of compromise than a framework name in a request header.
4. Hunt AWS CloudTrail activity
Look in CloudTrail management events, CloudTrail data events, GuardDuty, IAM records, VPC Flow Logs and relevant service logs.
Collect:
eventTime
eventSource
eventName
sourceIPAddress
userAgent
userIdentity.arn
userIdentity.accessKeyId
userIdentity.sessionContext
requestParameters
responseElements
errorCode
resources
Review sequences involving these APIs:
| Activity | APIs |
|---|---|
| Identity and permission discovery | GetCallerIdentity, ListRoles, ListUsers, GetRole, ListAttachedRolePolicies |
| Secrets access | ListSecrets, GetSecretValue |
| S3 discovery and data access | ListBuckets, ListObjectsV2, GetObject, PutObject, DeleteObject |
| Role assumption and compute activity | AssumeRole, DescribeInstances, RunInstances, CreateVolume, AttachVolume |
These are legitimate AWS operations. Their sequence, source and outcome determine whether they are suspicious.
Look for:
-
A workload identity authenticating from a new network and immediately enumerating permissions.
-
Numerous AccessDenied results followed by a successful GetSecretValue.
-
A newly assumed role starting EC2 instances or accessing resources outside its normal scope.
-
An identity reading large numbers of previously untouched S3 objects.
-
An unfamiliar identity performing storage reads, writes or deletions in a compressed time window.
For STS correlation, link the issued role session or access-key ID to subsequent requests. The original caller's ARN may change after role assumption.
S3 object-level operations such as GetObject require CloudTrail data-event collection, which is not enabled by default for ordinary trails. Without it, the absence of object-read events proves little.
5. Hunt Microsoft Entra ID, Azure and Microsoft Graph
Look in Entra service-principal sign-ins, AuditLogs, MicrosoftGraphActivityLogs, Defender XDR, Azure Activity Logs, Key Vault diagnostic logs and Storage resource logs.
Capture these fields where available:
Timestamp
TenantId
ApplicationId
ServicePrincipalId
IPAddress
ResourceId
Result / Status
CorrelationId
RequestUri
RequestMethod
ResponseStatusCode
Keep the global application ID distinct from the tenant-local service-principal object ID.
Microsoft's ARTEX/Storm-0940 investigation involved compromised ConnectWise application identities. Examples of affected application IDs include:
506e08f9-6e9a-4da4-942b-a099a03644ae
24205314-2b89-4e07-a2a4-579e898a5640
0de14c77-b9b0-4137-9360-34878a929827
These IDs identify legitimate application registrations involved in the reported campaign. Their presence alone does not indicate compromise.
Look for unusual service-principal sign-ins followed by Microsoft Graph activity, new application credentials, privilege changes or unexpected resource access. Review:
-
Service-principal authentication from previously unseen infrastructure.
-
New client secrets or certificates added to an application.
-
Privileged application-role assignments or consent changes.
-
Unusual enumeration of users, groups and directory permissions.
-
Bulk Key Vault secret reads across several vaults.
-
A sudden increase in Microsoft Graph requests or sensitive resource access.
Where service-principal sign-in records are aggregated, inspect the underlying event times. One visible row may represent several authentication attempts.
6. Monitor DNS, proxy and outbound network traffic
Look in DNS resolver logs, Zeek DNS/connection logs, firewall egress telemetry, secure web gateways, proxy logs and EDR network events.
Collect DNS questions and answers, source host, resolved addresses, destination IP/port, TLS SNI where available, request time, process identity and bytes transferred.
Contextual clues include:
-
webhook.site
-
New proxy domains.
-
Unfamiliar external AI gateways.
-
Unexpected tunneling or file-transfer services.
Google Threat Intelligence reported ARTEX-associated activity sending command output to rotating webhook.site endpoints.
The domain has legitimate uses. Correlate unexpected command execution with a new external HTTPS destination and unusual data transfer. Assess generic LLM-provider and webhook traffic against the workload and process involved.
7. Collect framework fingerprints
To discover servers running these frameworks, inspect public Internet scan records and your own asset inventory.
| Framework | Candidate fingerprint | Observation point |
|---|---|---|
| ARTEX | ARTEX-prefixed HTML title plus application assets | Exposed management frontend |
| Legacy ARTEX | /api/health with service: artex | Public API response |
| Reviewed ARTEX continuation | X-Artex-Auth: basic with basic_auth_required | Auth-gate response |
| AntiFox | AntiFox console title and service: antifox health response | Management service |
| PentAGI | PentAGI title and recognizable icon assets | Management frontend |
| PentAGI | /api/v1/info guest response | Public information endpoint |
| PentAGI | PentAGI CA or pentagi.local | Default generated TLS certificate |
Censys and Shodan can help discover exposed ARTEX and PentAGI instances. Title and API searches can identify AntiFox candidates.
A console fingerprint identifies a possible deployment. Establishing whether it attacked your organization requires evidence of that activity.
norma/0.4 is a shared HTTP-client marker. It cannot uniquely identify ARTEX or AntiFox activity. PentAGI/1.0 belongs to a PentAGI search client and may be absent from requests reaching a victim.
8. Prioritize correlated detections
| Priority | Detection | Correlation |
|---|---|---|
| P1 | Webshell or remote execution | Web request + file write + web-server child process |
| P1 | Cloud credential abuse | New identity origin + permission probing + successful secret access |
| P1 | Privileged application takeover | New app credentials/permissions + subsequent resource access |
| P1 | Unusual data access | Sensitive-object enumeration + bulk reads + unusual destination |
| P2 | Automated application reconnaissance | Endpoint breadth + high failure rate + follow-up successful action |
| P2 | Suspicious cloud retries | Denied API calls followed by successful related operations |
| P3 | Framework console fingerprint | Product-specific frontend/API/TLS evidence |
| P3 | Historical IOC match | Exact IP or user-agent clue requiring context |
Use these priorities as an operational starting point. Tune them against legitimate scanners, CI systems, administrators and expected application behavior.
9. Preserve evidence
For each strong detection, retain:
-
Raw events and normalized SIEM records.
-
The exact indicator and its original source.
-
First/last observation times.
-
Identity, session and request correlation IDs.
-
Related endpoint processes and files.
-
Cloud resources accessed and operation outcomes.
-
DNS, network-flow and proxy context.
-
Relevant benign automation baselines.
Preserve timestamps in UTC and retain the original event fields. Avoid logging secret values, authentication tokens or sensitive request bodies unnecessarily.
Build an identity-to-process-to-resource timeline that connects reconnaissance, authentication, execution, privilege changes and sensitive data access.
References
-
Microsoft: ARTEX and Storm-0940 / ConnectWise investigation: https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/
-
Google Threat Intelligence: ARTEX and Friends: https://cloud.google.com/blog/topics/threat-intelligence/artex-and-friends-enter-agentic-offensive-frameworks
-
Censys: ARTEX: Agentic Asbestos: https://censys.com/blog/artex-agentic-asbestos/
-
AntiFox-related intrusion reporting: https://www.bleepingcomputer.com/news/security/hackers-use-antifox-ai-agents-in-aws-credential-theft-attacks/
-
MalwareMustDie: ARTEX detection signatures and playbook: https://blog.malwaremustdie.org/2026/11/mmd-072-2026-artex-detection-signature.html
-
AWS: CloudTrail S3 data-event logging: https://docs.aws.amazon.com/AmazonS3/latest/userguide/enable-cloudtrail-logging-for-s3.html



