The original text is in EN, taken from the source site. Source ↗

Threat actors are weaponizing ARTEX, an autonomous penetration-testing framework, to accelerate network intrusions, escalate privileges, and steal sensitive information.

Google Threat Intelligence Group and Mandiant investigations described in the supplied report show attackers delegating substantial portions of post-compromise operations to coordinated AI agents, compressing execution sequences into seconds.

Observed since late August 2026, the activity spans financially motivated extortion groups and suspected cyberespionage operators.

It extends the transition documented in Google’s September report, “From Prompting to Autonomy,” from chatbot-assisted tasks toward agentic workflows that execute tools, interpret results, and troubleshoot operational failures.

Written in Go with an embedded Next.js console, ARTEX translates operator objectives into machine commands through six specialized sub-agents.

Investigators observed capabilities covering Active Directoryreconnaissance, cloud permission discovery, credential harvesting, privilege escalation, and custom webshell deployment.

Operators sometimes supplied compromised credentials and instructed the framework to continue the intrusion. ARTEX then tested access, generated scripts, and adapted to errors.

Individual execution bursts lasted approximately two to 25 seconds; one investigation recorded 27 events across 23 seconds while the framework probed AWS permissions.

Speed did not translate into precision. Against an internet-facing Weaver OA server, ARTEX used authenticated application APIs to issue SQL commands and attempted to convert database access into operating-system execution.

Its iterative approach generated 240 log events, including failures that may have been caused by competing sub-agents.

The framework attempted to abuse MySQL logging to place JSP webshells in web-accessible directories.

Compilation errors prompted repeated revisions until some files returned valid command output, illustrating adaptive exploitation alongside substantial forensic noise.

GTIG tracks, one ARTEX-associated campaign as UNC7228, with activity overlapping Microsoft’s China-linked Storm-0940 designation.

The supplied findings describe compromises affecting software, industrial automation, semiconductor, and telecommunications organizations across the United States and Europe.

Operators typically progressed from an initially compromised account to deep system access in approximately four hours, with one case taking just 15 minutes.

Wiz’s October research independently documented overlapping activity against IT and software vendors, including Okta, Microsoft Entra ID, AWS Cognito, and Microsoft 365 integrations.

Attack sequences combined application mapping, authentication probing, password spraying, account-recovery abuse, and cross-tenant access testing.

Wiz observed agents extracting API routes from downloaded source maps, then requesting those routes minutes later.

When blocked, activity shifted endpoints or infrastructure. Some sessions persisted for 24–30 hours, although most individual attempts failed.

Mandiant’s observations included database extraction, mass blob reads, extensive directory enumeration, and preparation of cloud storage for potential exfiltration.

Investigators also identified command output transmitted through rotating webhook.site endpoints.

Defenders should correlate rapid permission checks, repeated authorization failures, unexpected administrative accounts, webshell creation, unusual snapshot activity, and outbound webhook traffic.

Individual filenames or generic user agents are insufficient attribution signals; Wiz similarly cautions that many published user-agent indicators are noisy when hunted independently.

Recommended controls include isolated management planes, short-lived credentials, least privilege, segmentation, and behavioral detection capable of triggering rapid containment.

These measures address familiar weaknesses that agentic tools exploit at greater speed.

ARTEX’s developer announced its closure and discontinued maintenance on October 8 following malicious abuse. The announcement does not eliminate existing copies or forks.

Current evidence demonstrates successful, noisy automation across multiple intrusion stages not universal, stealthy, fully autonomous compromise.

The immediate challenge is shortening defensive response windows before automated experimentation finds exploitable access.

Stops Cyber threats before impact with 21 min faster MTTR. Integrate ANYRUN’s Sandbox in your SOC.