The Information reports hackers wired Claude, Grok and Kimi into an automated agent network that targeted more than 400 financial institutions
Hackers turned Claude and Grok into a network of AI agents that autonomously targeted more than 400 financial institutions and other financial targets, according to cybersecurity research shared with The Information. Read more:
Context
This is one of the clearest documented cases of criminals running several commercial AI models as a coordinated attack team. When one model refused a task, the group simply handed it to another, which shows how safety limits on a single model can be routed around. The findings come from one security firm and have not been independently confirmed.
The research comes from the security startup Straiker, which published its report on the CARLIY0S system on October 21, and The Information covered it with two posts on October 24. Straiker says the operators have targeted Latin American financial firms for more than a decade. The lead post was published at 10:00 AM Eastern on October 24.
Deep dive
The Information posted at 10:00 AM Eastern on October 24 that hackers turned Claude and Grok into a network of AI agents that autonomously targeted more than 400 financial institutions, citing cybersecurity research. A second post that afternoon added that the group used Grok to gather intelligence, Claude to plan attacks and Kimi for tasks Claude refused.
The research is from Straiker, which published its report on October 21 and says it worked from the group's own environment. It tracks the operators as Tr4nsHack Team, working under a front company called Hexagon Security Labs Mexico, and says they include members of CyberCartel, a group tracked since 2012. According to Straiker, the group built a system called CARLIY0S on top of PentAGI, an open source automated attack testing framework. A commander running Claude inside the Cursor editor handed work to Claude agents with fixed roles: STRIKER wrote exploits, GHOST did reconnaissance, FABLE looked for unusual angles, and MAXIMUS arbitrated their debate and returned one decision. When fallback Kimi and Sonnet agents refused, the system switched to prewritten local code.
Straiker says five Grok based miners fed a stolen secrets database it calls FinancialSoft. A September 24 snapshot held 4,493 records, including 2,911 email addresses, 903 account numbers, API keys, passwords and credentials that could enable abuse of SPEI, Mexico's interbank payment system. The operators framed requests as authorized security testing to get past model safeguards. Straiker says Anthropic revoked the credential the group used, and xAI is looking into the Grok activity.
These are claims from one security vendor, which sells defenses against such attacks. Victims are anonymized, and the report describes targeting rather than confirmed breaches at all 400 organizations.
What to expect next
Watch whether Anthropic and xAI publish their own accounts of the accounts they banned, as Anthropic has done for earlier misuse cases. Financial regulators in Mexico and the wider region may also respond if the stolen payment credentials prove valid. For model makers, the case adds pressure to detect misuse across a whole workflow rather than one prompt at a time, since splitting work across models defeated refusals here. Independent confirmation of the scale would also matter, because the 400 figure so far comes only from Straiker.