-
Straiker says CyberCartel used AI from Anthropic, xAI and China’s Moonshot AI to target over 400 banks and financial firms.
-
The hackers tricked the AI by claiming to run approved security tests. When one tool refused, the system switched to another tool or to ready-made attack code.
-
Anthropic has cut off the group’s access to Claude, and xAI is investigating. Total losses and the number of successful breaches remain unknown.
Criminals have found a new way to scale up bank attacks. They put AI programs to work as a team. Straiker’s report says a cybercrime group called CyberCartel used this approach against more than 400 financial targets across Latin America.
The AI handled much of the planning and data theft with little human help. The case also shows how easily a believable story can talk an AI into helping hackers.
CyberCartel is Reportedly Behind the Attacks
CyberCartel is not a new name. Researchers say it has been active since 2012. Earlier reports from IBM and Darktrace described its attacks on bank customers, mostly in Mexico and Chile.
For years, the group relied on manual hacking and simple scripts. However, that pace was too slow. Straiker says human speed became a bottleneck, so the group built an AI system named CARLIY0S. Straiker says the group now works as the Tr4nsHack Team. It also operates through a front company called Hexagon Security Labs Mexico. The firm believes the group has ties to Mexican cartels, but it calls that link presumed.
How the Attackers’ AI Team Worked
Straiker published its findings on October 21. The firm says it saw the group’s own files, including scripts, settings and a dashboard of stolen data. That gave researchers a rare look inside a criminal AI operation.
The group gave each AI model a job. Four Claude-based programs weighed attack options and argued over targets. One acted as the judge and made the final call. Another Claude model, running inside the coding tool Cursor, served as the commander. Meanwhile, five Grok programs gathered passwords and other secrets. Kimi, the model from Moonshot AI, served as a backup.
How They Got Past Safety Rules
AI companies build rules that block help with hacking. The group dodged those rules with a simple lie. It told the models the work was an approved security test or a school exercise. Straiker says the false stories worked repeatedly.
Still, some requests got refused. The system did not stop when that happened. Instead, it passed the job to another model. If the backups refused too, it ran a stored file of ready-made attack code.
Straiker says it knows of no earlier report of that fallback. The group also told its AI never to write the Tr4nsHack name. As a result, the output was harder to trace back to the group.
What Straiker Found in the Stolen Data
On September 24, Straiker reached a copy of the group’s stolen-data list. It held 4,493 records. They included bank account details, passwords and access keys. Those keys let programs connect to other systems.
The list held mostly identity and account data, including 2,911 email addresses and 903 account numbers. Straiker says some of the material could let criminals abuse SPEI, Mexico’s bank-to-bank payment network.
The group also tracked more than 200 security flaws in its targets’ systems, including some previously unknown ones. Straiker removed victim names to protect them, so the banks remain unidentified.
The Companies’ Response
Anthropic says it investigated after Straiker’s report and blocked the group from using Claude. Straiker adds that Anthropic revoked the access key the group used to reach Claude. Moreover, Anthropic’s recent threat report suggests the company is aware of the increased misuse of the models by attackers and is doing its best to counter it.
In addition, Straiker told xAI, which is now looking into the activity tied to the five Grok programs. Straiker also tried to contact Moonshot AI. Coverage so far has not included a reply from that company.
The Scope of the Attack Remains Obscure
Numbers need context here. Straiker says the group went after more than 400 targets. That does not mean 400 successful break-ins. The report does not say how many attacks worked or how much money was lost. Straiker says the AI appears to have stolen passwords and customer data and reached bank accounts.
One detail remains unclear. A settings file listed Anthropic’s newer Fable 5 model. However, Straiker says that does not prove the group ever used it. Also, Straiker sells AI security tools, so its findings come from a company with a stake in the topic. Still, the report rests on the group’s own files.
Implications of this Incident
Straiker’s main warning is about speed and scale. AI can help small crews hit more targets with fewer people. Because of that, banks cannot rely on slow, manual defenses. Straiker urges companies to test with AI, watch for odd activity, and lock down logins and access keys.
In a separate case, researchers reported an autonomous AI agent deploying custom ransomware to target AI models, underscoring the range of cyber threats emerging from AI-driven attacks.
Customers can take simple steps too. First, change your password. If you’ve ever reused the same password for multiple accounts, now is time to make some changes. Create unique and strong passwords for each account. Enable two-step login wherever it’s allowed.
Also, regularly monitor your bank statements so you can catch odd activity early. Doing these won’t stop a bank breach. However, they can limit the damage by preventing criminals from getting into your account and draining your funds.