Straiker: group tied to Mexico used Claude and Grok against banks
The security firm attributes to a group with alleged Mexican ties an AI system aimed at 400 financial institutions. No one else has confirmed it.
Cybersecurity firm Straiker says a group with alleged ties to Mexico used AI models from Anthropic and xAI to attack more than 400 banks, payment providers and other financial institutions in Latin America.
The report, published Oct. 21, describes a system called CARLIY0S. The firm tracks the group as AiPT-67. It presents itself as Tr4nsHack Team, with a front company called Hexagon Security Labs Mexico, and its members include operators of CyberCartel, a financial threat actor active since 2012. The Mexico link is a presumption: it rests on aliases and on documents the group claims.
CARLIY0S was built on PentAGI, a framework for offensive security testing, and combines Claude (Sonnet and Opus), Kimi and Grok. Claude-based roles debated each target, a referee decided, and five "Grok miners" gathered the data into a database called FinancialSoft. If one model refused a task, another served as backup.
What that database held on Sept. 24, and what the group had reportedly gathered, according to the firm:
- 4,493 records, including 2,911 emails and 903 CLABE account numbers, the standardized 18-digit Mexican bank key.
- More than 270 penetration-testing reports from the victims.
- More than 200 vulnerabilities, some of them zero-days with no published CVE.
- Loose credentials: 391 API keys, 67 unencrypted passwords, 43 database connection strings and 20 keys used to sign web sessions (JWT).
Anthropic revoked the compromised credential used to reach Claude, and xAI was notified and is investigating what the "Grok miners" did. Both responded quickly before publication, Straiker says. The two companies' accounts come only through the firm.
Straiker itself admits the limits. The analysis is based on what was found in the group's environment, did not verify every attack and does not identify each victim, who appear anonymized. The 400 are organizations targeted, not breached. No Mexican banks are named and no intrusions are confirmed.
On SPEI, Mexico's interbank payment system, the report suggests that credentials from a fintech provider could be used to abuse it. Straiker presents this as inference, not fact.
Darktrace had already described CyberCartel in January 2024 as a Latin American group with malware that steals information from users of banks and government services, mainly in Mexico and Chile. That analysis did not involve AI: the part about the models rests for now on a single report.
No affected institution or account is singled out. For companies, Straiker recommends continuously testing their AI agents and having a kill switch that revokes access and isolates agents without losing evidence.
