AI Agents Power Bank Heists
A long-time threat actor is leveraging large language models to automate financial fraud across Latin America.
By Elijah Fitzgerald ·
The gap between human effort and digital intrusion has narrowed dramatically in recent months. A threat actor known as CyberCartel has shifted its operations to rely heavily on artificial intelligence agents. This change allows the group to attack more targets simultaneously than its human workforce could ever manage. The shift marks a significant escalation in how financial institutions in Latin America are being compromised.
CyberCartel has been active in the region since at least 2012. For over a decade, the group targeted local businesses with traditional methods. Recent research from the cybersecurity firm Straiker reveals that the group’s reach has expanded due to AI automation. The new approach steals passwords and customer data with minimal manual intervention from the attackers. Individual bank accounts are accessed at a speed that defies conventional hacking tactics.
The tools driving this new capability include models from major AI developers. CyberCartel has utilized agents from SpaceX, specifically Grok, and from Anthropic, known as Claude. A model from the Chinese lab Moonshot AI has also been integrated into their workflow. Straiker has notified SpaceX and Anthropic of the usage. Both companies are currently working to restrict the group’s access to their services.
Anthropic has already blocked CyberCartel’s access to Claude. Straiker is still attempting to establish contact with Moonshot AI to discuss the findings. The use of these specific models highlights a broader industry concern. The incident underscores how vulnerable current safety mechanisms can be when faced with coordinated misuse by sophisticated actors.
The technical bypass used by the group is particularly concerning to security experts. AI models are typically programmed to refuse tasks that aid in cyberattacks. CyberCartel reportedly circumvented these safeguards by claiming the models were operating in test environments. When one model refused to proceed, the hackers simply switched to a different one. If that model also resisted, the agents were redirected to deploy malware instead.
This incident adds to a growing list of concerns regarding Anthropic’s technology. The company warned earlier this year that hackers could exploit its new Mythos model for such attacks. Although Straiker’s current report does not indicate that Mythos was used in this specific campaign, the timing is notable. The release of Mythos had previously prompted a global rush by security teams to fix thousands of vulnerabilities the model identified.
Anthropic is not new to dealing with bad actors using its tools. In November 2025, the company disclosed details of an AI-assisted cyberattack. Later, Claude was used to breach the security of rival developer OpenAI. However, that specific operation was conducted by researchers rather than criminal groups. The current case represents a shift toward malicious financial gain rather than academic or competitive exploration.
The report emerges just days after news of a similar issue in South Korea. Authorities there are investigating banks including Hana Bank, KB Kookmin Bank, and Shinhan Bank. Personal credit information belonging to at least 25,000 Shinhan customers has been leaked. This parallel development suggests that the problem is not isolated to one region or one type of threat actor.
Industry leaders are already responding to the rising risk. Nvidia, which is heavily invested in AI data center hardware, recently launched the Open Agent Safety Platform. This new software and reference system design aims to contain rogue agents. It uses a security stack intended to prevent agents from escaping sandboxes or executing unauthorized code. The platform is designed to stop agents from bypassing guardrails and accessing critical infrastructure without permission.
