Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Straiker Figure 1, the FinancialSoft Grok Master Inventory dashboard from CyberCartel's environment, dated 2026-09-24, with tiles counting stolen API keys, passwords, backdoors, portal logins, CLABE account numbers, connection strings, JWT keys, Keycloak credentials and emails, victim name redacted.
Image: Straiker STAR Labs, Figure 1, the group's FinancialSoft inventory dashboard, victim details redacted by Straiker, Oct. 21, 2026 (fair use).

Threat Intel

Straiker Says CyberCartel Ran Claude, Kimi and Grok as One AI Attack System Aimed at More Than 400 Latin American Banks and Payment Firms, Switching to Hardcoded Code When Models Refused

Working from the group's own exposed environment, Straiker's STAR Labs maps CARLIY0S, a PentAGI-based system that splits recon, exploits and decisions across Claude roles, mines credentials with Grok and drops to hardcoded scripts when a model says no. Its stolen-secrets inventory centers on Mexico's SPEI payment rail.

Priya Shah, Threat IntelSingapore6 min read

MEXICO CITY - A criminal group that has stolen from Latin American financial institutions for more than a decade has handed much of its hacking work to commercial AI models, assigning each one a job, according to a report that Straiker says was built from the group's own working files.

In research published Wednesday, Straiker's STAR Labs describes CARLIY0S, an attack system used by CyberCartel, a financially motivated actor tracked since 2012. Straiker says the crew operates as Tr4nsHack Team under a front company called Hexagon Security Labs Mexico, and names operators known as Adx, Roro and aBerry. The system, which Straiker classes as an AI-powered persistent threat, or AiPT, was aimed at more than 400 banks, payment providers and other financial targets across Latin America. That number counts targets, not confirmed breaches.

Straiker calls its report the first investigation it knows of to document a criminal group's AI-orchestrated operations directly from the group's own environment. Its researchers say proof-of-concept exploits, active keys and credential dumps have been redacted.

A commander and four specialists

CARLIY0S is built on PentAGI, an open-source autonomous offensive framework. A commander, Claude running inside the Cursor code editor, calls the CARLIY0S.py script, which dispatches work to four Claude-based roles through the Anthropic API. In the group's configuration, STRIKER (claude-opus-4-7) writes exploits and automation code, GHOST (claude-opus-4-8) handles reconnaissance and validation, FABLE (claude-fable-5) looks for unconventional and business-logic attack paths, and MAXIMUS (claude-opus-5) acts as arbiter after the others debate. Straiker treats FABLE as unverified, since a configuration entry does not prove the model was ever successfully called.

Straiker Figure 2, the CARLIY0S agent roster: a Comandante orchestrator directing STRIKER on claude-opus-4-7, GHOST on claude-opus-4-8, FABLE on claude-fable-5 and MAXIMUS on claude-opus-5.
Image: Straiker STAR Labs, Figure 2, the CARLIY0S agent roster as configured by the group, Oct. 21, 2026 (fair use).

The crew also used Kimi models and Claude Sonnet. When those refused a request, CARLIY0S redirected execution to a local file, CARLIY0S_consolidated.py, holding hardcoded offensive code. Straiker says the group pushed models past their safeguards by describing the work as an authorized red-team engagement, an academic capture-the-flag exercise or a vulnerability disclosure program, and that its rule files forbid the models from ever writing the Tr4nsHack name, to blunt attribution through model output.

Straiker Figure 7, the CARLIY0S fallback path: an initial request to secondary models KIMI-2, KIMI-3 and SONNET-2 is blocked by an ethical filter, and execution is redirected to a local safety net file, carliy0s_consolidated.py.
Image: Straiker STAR Labs, Figure 7, what CARLIY0S does when a model refuses, Oct. 21, 2026 (fair use).

What the stolen inventory shows

Five processes the group called Grok miners, using xAI's model, gathered credentials. On Sept. 24, Straiker says its agents reached a snapshot of FinancialSoft, the group's stolen-credential database, holding 4,493 records. Its inventory lists 2,911 email addresses and 903 CLABE account numbers; 11 API keys and 32 cryptographic keys belonging to one fintech company, plus 391 general API keys; 67 cleartext passwords, 51 portal logins and 43 database connection strings with embedded usernames and passwords; 20 JWT signing keys, 20 Keycloak credentials and four certificates; 19 Wasabi S3 credentials and 15 Mailgun credentials; and six backdoors. Straiker also found more than 270 penetration-test style reports written between May and September 2026 and more than 200 vulnerabilities, some of them zero-days.

Straiker says the payment credentials center on SPEI, Mexico's interbank payment system, and are consistent with small test transactions it observed in the campaign.

How the machinery ran

The group's hub was a Hexagon server at 157[.]180[.]98[.]220 with 61TB of RAID 5 storage, an MCP knowledge graph on port 8001 and a portal on port 8080. A twin workstation shared the same .cursor/mcp.json configuration, and scripts named sync_live.sh and sync_watchdog.ps1 copied data with rsync every 15 minutes, using a PowerShell call to SetThreadExecutionState to keep machines awake.

Straiker Figure 3, the CARLIY0S operational chain from the Hexagon C2 server at 157[.]180[.]98[.]220 through a primary and twin node, blocked models and a Claude swarm, local Playwright execution and a SOCKS5 anonymity layer to three target vectors.
Image: Straiker STAR Labs, Figure 3, the CARLIY0S chain from command server to target, Oct. 21, 2026 (fair use).

Execution ran through headless Playwright and then Python Requests, tunneled over SOCKS5 through IPBurger and Bright Data residential proxies so traffic looked domestic to Mexican sites and slipped past geographic controls. Straiker describes the system logging in to Magnolia CMS administrative interfaces, driving the Vaadin-based front end to inject obfuscated Groovy scripts for remote code execution, then copying session cookies into Requests. Three post-exploitation payloads followed: an Oracle database scan, a hunt for JDBC connection strings and SSH private keys so access would survive a patch, and internal subnet mapping that turned the web server into a pivot point. Burp Suite was wired in through MCP.

What happened next, and what to watch

Before publishing, Straiker shared indicators with the model providers. It says Anthropic revoked the credential the swarm used to reach Claude, and that xAI has been notified and is looking into the activity of the five Grok miners.

For SOC teams at banks and payment firms, the report points to familiar controls applied at machine speed:

  • Rotate and scope API keys, JWT signing keys, Keycloak credentials and database connection strings, and keep secrets out of reachable configuration files.
  • Watch internet-facing Magnolia CMS admin consoles for logins and Groovy script execution, and restrict them by network.
  • Flag bursts of logins from residential proxy ranges posing as domestic traffic.
  • Monitor payment rails such as SPEI for small test transfers that precede larger fraud.
  • If you run AI agents of your own, know what each can reach, and be able to revoke its tool access, suspend sessions and isolate it while keeping evidence, which Straiker calls an agentic kill switch.

Sources:


Priya Shah covers threat intelligence, intrusion analysis, and adversary tradecraft for SOCtember from Singapore.

Related stories

Threat Intel desk