Grok and Claude among AI agents used to automate attacks on Latin American banks, researchers say

Straiker reports CyberCartel's test-environment ruse; Anthropic blocks Claude access

By LineZotpaper

Published23 October 2026

Read Time2 min

A hacking group with a long history of targeting Latin American businesses has been using AI agents from Anthropic, SpaceX and Chinese lab Moonshot AI to automate attacks on banks and mobile payment apps, according to research from cybersecurity startup Straiker.

Cybersecurity startup Straiker has found that CyberCartel, a group active since at least 2012, is using AI agents such as Anthropic's Claude and SpaceX's Grok to automate attacks on Latin American financial institutions. The research, shared with The Information, says the group has stolen passwords and customer data and accessed individuals' bank accounts with limited human input.

According to Straiker, the group bypassed safety protections by telling the models they were operating in test environments. When one model refused, the hackers switched to another, and if that also refused, the AI agents were tasked with deploying malware instead. The approach has let the group attack more businesses at a speed that was not possible with human-only hacking.

Straiker shared its findings with SpaceX and Anthropic. Both firms are working to lock CyberCartel out of their services, and Anthropic has already blocked access to Claude. Straiker said it is still trying to contact Moonshot AI. Notably, the group does not appear to be using Anthropic's Mythos, a model whose release prompted warnings that hackers could use it for attacks like these.

The report follows earlier disclosures of AI misuse. Anthropic shared details of its AI being used in a cyberattack in November 2025, and researchers later used Claude to break the security of OpenAI, an Anthropic competitor. It also comes days after news that South Korean authorities are investigating similar attacks on local banks, with at least 25,000 customers of Shinhan Bank having their personal credit information leaked.

In response to the broader threat, Nvidia has launched the Open Agent Safety Platform, a hardware and software security stack designed to keep rogue AI agents from escaping their sandboxes, executing unauthorized code, gaining unauthorized access to critical infrastructure, or bypassing guardrails.

Analysis

Why This Matters

AI agents let a small group attack many banks and payment apps at once, making cybercrime cheaper and faster.
Simple social engineering is defeating commercial AI safety guardrails, raising questions about how models are protected.
Financial institutions in Latin America are the early targets, and the same techniques may spread elsewhere.

Background

Commercial AI models are built with safeguards intended to stop them being used for harmful purposes, but researchers have repeatedly shown that these can be circumvented. The Straiker findings describe a pattern in which attackers treat refusal from one model as a minor obstacle, switching between Grok, Claude and Moonshot AI's agent until one complies. The South Korean investigation into similar bank attacks suggests the problem is not limited to one region. Periodic disclosures of AI-assisted intrusions, including Anthropic's November 2025 report, keep the issue in front of security teams.

Key Perspectives

Anthropic and SpaceX: Both companies were told of the research and are working to block CyberCartel. Anthropic has already prevented the group from using Claude, according to Straiker. Security researchers: Straiker argues current protections are surprisingly easy to bypass and that the automation lets attackers scale up their operations far beyond human capacity. The attackers: CyberCartel, according to the research, relies on the test-environment ruse and on model-switching, and falls back to malware deployment when models refuse. Critics and skeptics: Security teams may question whether voluntary cooperation from AI labs is enough, pointing to Moonshot AI not yet being contacted and the continued availability of multiple alternate models. Some look to technical fixes such as Nvidia's Open Agent Safety Platform.

What to Watch

Whether Anthropic and SpaceX can keep CyberCartel locked out, and whether Moonshot AI responds to Straiker.
The outcome of the South Korean investigation into similar attacks on Hana Bank, KB Kookmin Bank and Shinhan Bank.
Adoption of agent-security hardware and software such as Nvidia's Open Agent Safety Platform.

Sources

Claude and Grok among AI agents used to automate attacks on Latin American institutions · Latest from Tom's Hardware

Checked

12 claims in this story checked against 1 source: 9 supported.

Show each claim

SupportedGrok and Claude were among the AI agents CyberCartel used to automate attacks on Latin American banks.Latest from Tom's Hardware: “CyberCartel, a hacking group with a long history of targeting Latin American businesses, has been using AI agents, including SpaceX's Grok and Anthropic's Claude, to automate attacks on financial institutions such as banks and mobile payment apps.”
SupportedCyberCartel has been active since at least 2012.Latest from Tom's Hardware: “While CyberCartel has been an active threat since at least 2012, its new leveraging of the power of AI has leveled up its attacks in recent months.”
SupportedCyberCartel stole passwords and customer data and accessed individuals’ bank accounts with little human input.Latest from Tom's Hardware: “Passwords and customer data have been stolen, while individuals' bank accounts have also been accessed with little input from criminals”
Partly supportedCyberCartel bypassed model protections by telling models they were operating in test environments.Latest from Tom's Hardware: “In this case, CyberCartel reportedly simply told the models that they were being used in test environments.”
SupportedWhen one model refused, the hackers switched models; if that model also refused, the agents were tasked with deploying malware.Latest from Tom's Hardware: “If one model refused to proceed, the hackers used a different model. If that also refused, the AI agents would be tasked with deploying malware instead.”
SupportedThe approach enabled attacks on more businesses at a speed not previously possible with human-powered hacking.Latest from Tom's Hardware: “Now, the research shows that the group can attack more businesses at a speed previously not possible when using human-powered hacks.”
SupportedStraiker shared its research with SpaceX and Anthropic; both are working to block CyberCartel, and Anthropic has already blocked Claude access.Latest from Tom's Hardware: “Straiker shared its research with both SpaceX and Anthropic, and both firms are working to lock CyberCartel out of their services. The latter has already been able to block access to Claude.”
SupportedCyberCartel also used an agent from Moonshot AI, and Straiker is still trying to contact the company.Latest from Tom's Hardware: “Chinese lab Moonshot AI's agent has also been used by the hacking group, but Straiker is still trying to get in touch with the company.”
SupportedCyberCartel does not appear to be using Mythos, and Anthropic warned hackers could use the model for attacks like these.Latest from Tom's Hardware: “Notably, CyberCartel doesn't appear to be using Anthropic's Mythos for its recent attacks. The company caused a stir earlier this year when it announced the new model, warning that hackers could use it and models like it for attacks like those carried out by CyberCartel.”
SupportedAnthropic reported AI being used in a cyberattack in November 2025, and researchers later used Claude to break OpenAI’s security.Latest from Tom's Hardware: “The company shared details of its AI being used to carry out a cyberattack in November 2025. Later, Claude was used to break the security of OpenAI, an Anthropic competitor in the AI space. That hack was carried out by researchers rather than criminals, at least.”
Partly supportedSouth Korean authorities are investigating similar attacks on local banks, with at least 25,000 Shinhan Bank customers’ personal credit information leaked.Latest from Tom's Hardware: “South Korean authorities are investigating a similar issue affecting local banks. Hana Bank, KB Kookmin Bank, and Shinhan Bank are among the South Korean banks affected, with at least 25,000 Shinhan Bank customers having their personal credit information leaked.”
Partly supportedNvidia launched a security platform intended to prevent rogue agents from escaping sandboxes, executing unauthorized code, gaining unauthorized access to critical infrastructure, or bypassing guardrails.Latest from Tom's Hardware: “The software platform and reference system design were developed to secure rogue AI agents via a new hardware and software security stack. It's hoped that it will go some way toward preventing agents from escaping their sandboxes, executing unauthorized code, gaining unauthorized access to critical infrastructure, or bypassing guardrails.”