What happened
Straiker, a cybersecurity startup, reported that the hacking group CyberCartel has begun using AI agents—including Anthropic’s Claude and SpaceX’s Grok—to automate credential‑stealing attacks on financial institutions in Latin America. The group allegedly accessed passwords, customer records, and even individual bank accounts from over 400 targets, ranging from traditional banks to mobile payment apps. Straiker’s findings were shared with both Anthropic and SpaceX, which have since taken steps to block the gang’s access to their services. The report also mentions that a Chinese lab, Moonshot AI, provided an additional agent that CyberCartel used, though Straiker has not confirmed contact with that firm.
Straiker’s research, as reported by Tom’s Hardware, documents a shift in CyberCartel’s tactics from traditional phishing and credential‑stuffing to AI‑driven automation. The group reportedly feeds prompts to Claude and Grok that instruct the agents to locate, extract, and exfiltrate login credentials from targeted services.
The stolen data includes usernames, passwords, and in some cases, one‑time passcodes, enabling direct unauthorized access to customer bank accounts. Straiker says the attacks were carried out with minimal human oversight, relying on the agents to navigate login pages and bypass basic security checks.
Anthropic confirmed that it has blocked the offending accounts from accessing Claude, while SpaceX said it is working with Straiker to harden Grok against similar abuse. No public comment was obtained from Moonshot AI.
Source details: tomshardware.com ↗
Why it matters
The incident shows that commercially available AI agents can be weaponized to scale cyber‑crime, dramatically reducing the manual effort required to breach large numbers of institutions. If attackers can automate credential harvesting at this speed, the threat surface for banks and fintech services widens, especially in regions with weaker security postures. Anthropic’s rapid response—blocking Claude access—highlights the need for AI providers to embed robust abuse‑prevention controls. The episode also raises broader concerns about the dual‑use nature of powerful language models and the responsibilities of AI developers to monitor and mitigate malicious use.
AI agents can process large numbers of login portals in parallel, a capability that traditional hacking tools lack. This scalability could lead to a surge in successful breaches if not curbed.
The incident underscores a gap in current AI service terms of use and monitoring. Providers may need to implement real‑time abuse detection, credential‑leak prevention, and stricter API access controls.
Financial institutions in Latin America, many of which operate with limited cybersecurity budgets, may become prime targets unless they adopt stronger multi‑factor authentication and anomaly detection.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Why can ethical evaluation not be reduced to one model score?
What to watch next
Watch for further statements from Anthropic and SpaceX on additional safeguards they plan to deploy for their agents, such as stricter authentication, usage‑monitoring, or rate‑limiting. Monitor whether other AI vendors experience similar abuse and how industry groups respond with best‑practice guidelines. Regulators in Latin America may also consider new cybersecurity mandates for financial institutions in light of AI‑enabled attacks.
Anthropic’s roadmap for Claude’s abuse‑prevention features, including any changes to the “allow” or usage quotas.
SpaceX’s upcoming security updates for Grok, especially any sandboxing or credential‑filtering mechanisms.
Potential legislative or regulatory actions in Brazil, Mexico, and Argentina aimed at mandating AI‑risk assessments for fintech providers.
Industry‑wide collaborations, such as information‑sharing coalitions, that could help surface AI‑driven threat intelligence faster.