CyberCartel used frontier AI models to automate attacks on 400 banks

The Information reports that a hacking group called CyberCartel used AI agents powered by frontier models from Anthropic, SpaceX and the Chinese lab Moonshot to automate attacks against more than 400 Latin American financial targets. The research came from the cybersecurity company Straiker and was shared exclusively with The Information. The piece is paywalled and Straiker has not published the investigation separately on its own site.

Those are the three facts the outlet reports, and the ones a reader can take on at this point: the actor and the scale, the vendors whose models were used, and that the research sits with Straiker and The Information. The Information's reporters, Aaron Holmes and Cory Weinberg, say this is one of the first detailed accounts of criminals successfully stringing a frontier model into every step of an attack on financial institutions.

What this is not

CyberCartel is a known Latin American threat actor. Darktrace published an investigation into the same name in January 2024 (phishing, malvertising, fake Mexican government portals) and no AI model is named in that write-up. The research the current article covers is a separate, 2026 campaign on the same actor, and the news is the AI-automation part.

A separate story that is easy to confuse this with: CrowdStrike's investigation into attacks on South Korean banks, published on 8 October, which named a Chinese threat actor using the open-source agent ARTEX together with Claude Code. Different victim country, different researcher, different attack toolchain. The CyberCartel reporting is new work.

The usage pattern the model providers now face

An attack that is one step is an input a filter can catch. An attack that is a hundred steps, each a plausible task when read on its own (summarise this document, write a login handler, format this list of numbers), is a session, and the signal that something is wrong lives at the account level rather than the prompt level. That is the shape the Information piece describes, which is the shape every vendor whose model is named will now have to reckon with.

For a team shipping anything a frontier model powers, this is the kind of event that moves the conversation on provider abuse detection past keyword filters and towards behaviour-level signals: unusual tool chains, out-of-character task mixes, accounts whose cadence changes overnight. None of that is new in security, and none of it is standard practice for the model APIs yet.

Source

Hackers Used Claude and Grok to Fully Automate Attacks on Banks, The Information, by Aaron Holmes and Cory Weinberg. The article is paywalled.

Context: Darktrace's older CyberCartel investigation and CrowdStrike's recent ARTEX report.