A hacker group called CyberCartel used a network of AI agents built on models from Anthropic, SpaceX and China's Moonshot AI to automate attacks on more than 400 Latin American financial targets, The Information reported on Wednesday, citing an investigation by cybersecurity startup Straiker.
The report describes a campaign that went beyond AI-assisted hacking. According to Straiker's findings as relayed by The Information, the agents operated with little human input: they autonomously stole passwords and customer data and accessed bank accounts on their own.
A Multi-Model Attack Stack
What distinguishes the operation, per the reporting, is the breadth of the model stack behind it. The group drew on Anthropic's Claude and SpaceX's Grok — the latter now part of Elon Musk's SpaceX following the aerospace company's February 2026 acquisition of xAI — alongside Kimi, developed by China's Moonshot AI.
Using several rival models at once gives an operation both redundancy and specialization: different models have different refusal behaviors, different strengths in code generation and different cost structures. Security researchers have long warned that mixing providers can help attack operations keep running when one lab's safety systems cut off access.
The scale is the other headline figure. More than 400 Latin American financial targets were hit by the automated agent network, according to The Information. The report did not identify the affected institutions, and details about total financial losses were not immediately available.
From AI-Assisted to AI-Automated
Security teams have spent two years tracking criminals who use chatbots to write phishing lures, debug malware or translate lures into new languages. The CyberCartel campaign, as described, marks a step further: agents that chain those steps together and execute them without a human driving each action.
That shift changes the economics of cybercrime. A human-operated intrusion scales with headcount; an agent-driven one scales with compute and API access. Straiker's account of agents stealing credentials, exfiltrating customer data and accessing accounts autonomously suggests the group treated commercial AI models as operational infrastructure rather than as an occasional helper.
The financial sector has been the proving ground for this trend all year. Earlier in October, reports surfaced that AI tools — including agents built on Claude — were used in cyberattacks on South Korean banks, in a campaign that BleepingComputer linked to a system called ARTEX AI and that Korean outlet Yonhap reported might involve a Chinese-speaking hacker. The South Korean episode and the newly reported Latin American campaign share a pattern: mid-sized financial institutions, AI-driven tooling, and cross-border investigation teams piecing together what happened after the fact.
The Defense Problem
The report lands at a delicate moment for the AI labs whose models were allegedly repurposed. Anthropic in particular has built much of its enterprise pitch on being the safety-focused provider, publishing disclosures when it detects and bans misuse of Claude for cyberattacks. A case in which Claude models sat inside an active attack operation — alongside competitors' models — underscores how little control any provider has over what happens to its outputs once an account, or a stolen API key, is in a criminal's hands.
It also sharpens the debate over model-level cyber capabilities. Frontier models now write, test and debug exploit code well enough that both offensive researchers and defenders use them daily. Providers have responded with tiered restrictions on cyber-related requests and monitoring systems that flag abuse patterns. An operation like CyberCartel's, spread across three providers, tests whether those controls matter when attackers can simply switch vendors mid-campaign.
Straiker's investigation, as reported by The Information, did not detail which banks were affected, how the group initially obtained access to the models, or how much money was taken. Those open questions matter for assessing whether the campaign represents industrial-scale theft or a large number of smaller intrusions, and they remain unanswered in the initial reporting.
What Comes Next
Expect three threads to develop. First, financial regulators in Latin America will face pressure to disclose which institutions were affected, once national CERT teams confirm the intrusions. Second, the labs named in the report — Anthropic, SpaceX and Moonshot AI — will likely face questions about what their abuse-monitoring systems saw and when. Third, security vendors will push agent-behavior monitoring as the countermeasure, arguing that the response to autonomous attacks has to be autonomous detection.
The campaign is a reminder that the same agent frameworks being marketed to enterprises for legitimate automation are already being run in anger against critical infrastructure. The defenders' advantage — visibility, telemetry, and the ability to coordinate across institutions — only pays off if banks treat AI-driven intrusion as an operational assumption rather than a hypothetical. For now, the CyberCartel report stands as one of the clearest documented cases of AI agents executing a financial intrusion campaign end to end, and the industry's response over the coming weeks will show whether the lesson sticks.
Stay Ahead of AI
Track AI security incidents and industry response with AI Buzz Wire's independent coverage of the latest AI developments.
Read more AI news →
