According to Perera's account, the attackers allegedly bypassed the AI systems' safety restrictions by presenting themselves as an authorised red team. Red teaming involves simulating cyberattacks to identify vulnerabilities in systems and strengthen security.
READ THIS: OpenAI's AI agents disabled safeguards meant to monitor them. One allegedly lied about it
When the AI models reportedly refused to carry out certain requests, the attackers allegedly switched to pre-written attack code. This approach suggests that the operation may have combined AI-assisted activities with conventional cyberattack techniques, potentially allowing the criminals to automate parts of their work while bypassing some safeguards.
JUST IN: A crime group used Claude, Grok and Kimi AI agents running in parallel to attack more than 400 banks and financial firms across Latin America, per security firm Straiker. They got past the guardrails by posing as an authorised red team. When models refused, it switched to pre-written attack code. The Ai agents appear to have stolen passwords and customer data and accessed bank accounts with minimal human input. Anthropic has cut off its access, and xAI is investigating.
— Shanaka Anslem Perera ⚡ (@shanaka86) October 22, 2026
Passwords and customer data allegedly stolen
The AI agents allegedly collected passwords and customer information and gained access to bank accounts with limited human intervention. Straiker's investigation reportedly uncovered a database containing thousands of records, including account details, credentials and other sensitive information. However, the names of the targeted financial institutions have not been publicly disclosed.
The reported figure of more than 400 refers to targeted financial organisations and should not be interpreted as confirmation that every institution was successfully breached. The findings are based on Straiker's investigation and have not been independently verified across all the alleged targets.
Anthropic blocks access, xAI investigates
Following the discovery, Anthropic, the company behind Claude, reportedly investigated the activity and revoked the compromised credential used by the attackers to access its systems. According to Straiker, xAI was also notified and began investigating activity associated with the use of Grok.
ALSO READ: 'There won't be money anymore': Yuval Noah Harari backs Elon Musk's prediction, warns of AI takeover
The incident underscores the cybersecurity risks associated with increasingly autonomous AI systems. While these tools can help organisations identify vulnerabilities and strengthen their digital defences, criminals could also exploit their ability to automate reconnaissance, process information and coordinate complex operations.
Security experts face the challenge of detecting attacks that can operate across multiple targets at machine speed. The reported campaign also raises questions about whether existing safeguards are sufficient to prevent AI models from being manipulated into supporting malicious activities.