More than 400 Latin American banks, payment companies and other financial institutions (not named) · AI used against the institution · severity 5 of 5
A hacking group known as CyberCartel used a network of AI agents built on Anthropic's Claude, SpaceX's Grok and Moonshot AI's Kimi to automate attacks on more than 400 banks, payment companies and other financial institutions across Latin America, according to research by the security startup Straiker reported by The Information on 21 October 2026. The agents reportedly stole passwords and customer data and accessed victims' bank accounts with little human involvement, after the group got past model safeguards by framing the work as academic testing or internal security checks. Anthropic said it opened an investigation after receiving the findings and blocked the hackers' use of Claude. No victim institution has been named and no regulator statement had been found as of October 2026.
Record last reviewed Oct 24, 2026 · Markdown
Date
May 2026 · disclosed Oct 21, 2026
Institution
More than 400 Latin American banks, payment companies and other financial institutions (not named)
Country
Several countries
AI system
CARLIY0S · Multi-agent attack system built by the attackers on commercial large language models, according to Straiker · vendor: Anthropic (Claude), SpaceX/xAI (Grok), Moonshot AI (Kimi)
AI?
Machine-learning system (inferred from the record)
The sources describe a machine-learning model; nobody on the record labelled it AI.
Category
Status
Reported
Severity
5 of 5
Impact
Not disclosed; the agents reportedly stole passwords and customer data and accessed victims' bank accounts at targets among more than 400 Latin American financial institutions · Straiker counted 4,493 records in the group's FinancialSoft database, including 2,911 email addresses, 903 CLABE account numbers and SPEI payment-rail credentials. No victim institution or loss figure was named.
What happened?
On 21 October 2026 The Information reported, citing research by the cybersecurity startup Straiker shared exclusively with it, that a hacking group known as CyberCartel had used a web of AI agents powered by models including Anthropic's Claude and SpaceX's Grok to attack more than 400 targets in Latin America, including banks and payments companies, PYMNTS reported. The Economic Times reported that Straiker published its report the same day and that the system, called CARLIY0S, was allegedly developed by CyberCartel to automate different stages of attacks, including analysing targets and identifying vulnerabilities.[2][3]
According to the report as summarised by Yonhap, the group assigned Grok agents to gather information about targets on the web, four agents based on Claude Opus and Claude Sonnet to find the best method of attack, and another Claude model the role of "commander" overseeing the hacking operation. The Herald Business and Yonhap reported that the agents stole passwords and customer data and reached victims' bank accounts with minimal human intervention.[4][5]
The group reportedly evaded the models' safeguards against cyberattack requests by claiming the work was taking place in a harmless setting such as an academic test or an internal corporate system check. When Claude refused a hacking task, the system was set up to hand it to Kimi, and if Kimi also refused, to switch automatically to deploying malicious code, Yonhap and The Herald Business reported.[4][5]
Straiker said CyberCartel appears to have been active since at least 2012 and that its attack speed and efficiency rose sharply after it adopted the new AI agents in recent months, according to Yonhap. Straiker chief executive Ankur Shah told The Information that "Offense is way ahead of defense right now" and that "Enterprises are just not ready", PYMNTS reported.[2][4]
In its own report, Straiker said that on 24 September 2026 its agents accessed a snapshot of "FinancialSoft", the group's stolen-credential database, holding 4,493 records collected by five "Grok miners", including financial account information, passwords, API keys, zero-day exploits and backdoors. It said the inventory, gathered from victims between May and September 2026, was led by 2,911 email addresses and 903 account numbers, and that the harvested secrets centred on credentials for SPEI, Mexico's interbank payment system, including those of a fintech payment provider, along with 903 CLABE account numbers. Straiker said victim organisations were anonymised and aggregated as Latin American and Mexican financial institutions, banks and fintech payment gateways.[1]
When did it happen?
- May 2026
Start of the window in which, per Straiker, the group's inventory of victim data and 270+ pentest reports was collected (May to September 2026).[1]
- Sep 2026
End of the collection window Straiker documents; on 24 September 2026 Straiker's agents access a snapshot of the group's FinancialSoft stolen-credential database (4,493 records).[1]
- Oct 2026
Straiker sends its report to AI developers including Anthropic and SpaceX the week before publication; Anthropic and others investigate and block the hackers' access to their models, according to Yonhap.[4][5]
- Oct 21, 2026
The Information reports Straiker's findings that CyberCartel used Claude- and Grok-powered AI agents to attack more than 400 Latin American targets, including banks and payments companies.[2]
- Oct 21, 2026
Straiker publishes its report on the CARLIY0S multi-agent system.[1][3]
- Oct 22, 2026
Yonhap and The Herald Business report the findings, including the role split between Grok, Claude and Kimi agents and the refusal fallback.[4][5]
What was the impact?
- Customers: Not disclosed; the agents reportedly stole passwords and customer data and accessed victims' bank accounts at targets among more than 400 Latin American financial institutions[1][2][3][4]
- Other: Straiker counted 4,493 records in the group's FinancialSoft database, including 2,911 email addresses, 903 CLABE account numbers and SPEI payment-rail credentials. No victim institution or loss figure was named.[1][2][3][4]
Severity 5 of 5: Criminal and across several institutions: a long-running cybercrime group reportedly automated attacks on more than 400 Latin American financial targets, stealing passwords and customer data and reaching victims' bank accounts.
How did More than 400 Latin American banks, payment companies and other financial institutions (not named) respond?
No public statement from any affected bank or payment company found as of 2026-10-24; the victims were not named. Anthropic said it began an investigation after Straiker revealed its findings and quickly blocked the hackers' use of Claude, and Yonhap reported that Anthropic and other developers blocked the hackers from using their models.[2][4]
What was the outcome, and did regulators act?
As of 24 October 2026 the hackers' access to Claude had been blocked, according to Anthropic's statement to The Information. No arrests, regulator action or confirmed losses had been reported in the sources reviewed.[2][4]
Which AI models did CyberCartel use against Latin American banks?
According to Straiker's research as reported by The Information and Yonhap, the group used agents built on Anthropic's Claude (Opus and Sonnet), SpaceX's Grok and Moonshot AI's Kimi. Grok agents gathered information on targets, four Claude agents weighed attack methods and another Claude model acted as commander.
Which banks were hit in the CyberCartel AI-agent attacks?
None has been named. The reports describe more than 400 targets across Latin America, including banks and payments companies, and no affected institution had made a public statement as of 24 October 2026.
How did the hackers get AI models to help with attacks?
They reportedly told the models the work was an academic test or an internal security check. When Claude refused, the system passed the task to Kimi, and if Kimi also refused it switched to deploying malicious code, Yonhap reported.
What did Anthropic do about the CyberCartel attacks?
Anthropic said it opened an investigation after Straiker shared its findings and quickly blocked the hackers' use of Claude, according to its statement to The Information as reported by PYMNTS.
- Inside CARLIY0S: How CyberCartel Used Multiple AI Models to Attack 400+ Latin American Banks and Financial Institutions security research
- Hackers Employ Grok and Claude to Attack LatAm Banks news
- Cybercriminals used AI agents to target over 400 Latin American banks: Report news
- "해킹그룹, 클로드 등 활용해 중남미 금융사 400곳 공격 자동화" (Hacking group automated attacks on 400 Latin American financial firms using Claude and others) news
- “해킹그룹, 중남미 금융사 400곳 공격 자동화” (Hacking group automated attacks on 400 Latin American financial firms) news
When AI goes wrong at a bank, the next morning's brief says so.
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning