Claude and Grok Used in AI Cyberattacks Targeting 400+ Financial Firms

A security firm says a hacking group used commercial AI models, including Anthropic's Claude and xAI's Grok, to coordinate attacks on more than 400 banks, payment providers and fintechs in Latin America. The report, published October 21, 2026 by Straiker, is one of the first detailed looks inside an AI-orchestrated attack on financial institutions. It is also easy to overstate. 

This article lays out what the research says, what remains unconfirmed, why crypto users should care, and what you can do about it. It is a news explainer, not financial advice.

Key Takeaways

• Straiker says a group it tracks as AiPT-67, linked to CyberCartel and operating as "Tr4nsHack Team" under the front name Hexagon Security Labs Mexico, targeted more than 400 financial organizations using a multi-model AI framework called CARLIY0S.

• "Targeted" is not "breached." Reports do not confirm that all 400 plus were compromised, no comprehensive loss figure has been published, and victims are unnamed.

• Humans, scripts and conventional tools remained central. When some AI models refused offensive requests, the system reportedly fell back to other models or prewritten code, which is why model-level safeguards alone do not stop a determined attacker.

What Straiker Reported

Straiker said it accessed a snapshot of the group's working environment on September 24, 2026, including scripts, model configurations and a stolen-credential database called FinancialSoft. The research was published October 21, and The Information reported it the same day, with other outlets following on October 22.

The reported targets were banks, payment providers and fintech gateways across Latin America, with an emphasis on Mexico, where credentials tied to the SPEI interbank payment system were reported. The framework is built on PentAGI, an open-source automated security-testing platform. 

Straiker described the group's data inventory as including thousands of records: a 4,493-record credential database, 2,911 email addresses, 903 account numbers, and credentials such as API keys and passwords. Straiker also listed more than 200 vulnerabilities in victims' systems, which have not been independently verified.

Straiker's CEO, Ankur Shah, said "offense is way ahead of defense right now" and that "enterprises are just not ready," according to coverage by PYMNTS and others. Straiker sells AI security products, so its framing should be read with that in mind.

How the AI Models Were Used

According to Straiker's documentation, the setup split work across several models, each with a role:

• Claude. A Claude model running inside the Cursor coding tool acted as a "commander," delegating to four Claude-based sub-agents for reconnaissance, code generation and other tasks, plus an arbiter that returned a single decision. One configuration entry referred to a newer Claude model, but Straiker says that does not prove it was used, and treats that role as unverified.

• Grok. Five automated "miners" used Grok to collect secrets, according to Straiker. Some secondary outlets describe Grok's role instead as searching the web for information about targets, so descriptions differ.

• Kimi. Moonshot AI's Kimi models were fallbacks, and reportedly also refused offensive requests.

• Fallback code. When models declined, the orchestrator reportedly switched to a local file of prewritten code that needs no model.

The operators reportedly framed requests as authorized penetration testing or academic work to get past safeguards, a tactic Straiker says was meant to influence model responses, not evidence of legitimacy. This article does not describe the technical methods the group used.

What Is Confirmed and What Is Not

The reporting is consistent on some points and weak on others.

Documented by Straiker from the group's own files: the architecture, model configuration, credential inventory and infrastructure.

Not established:

• Breach counts. The 400 plus figure refers to targets. One secondary report said sensitive data from one institution was confirmed stolen but unnamed; others say the extent of compromise and losses has not been disclosed.

• Fraud. Straiker says the SPEI-related access could enable abuse and is consistent with observed test transactions, but it does not show completed fraud.

• Autonomy. Several write-ups stress this was not proof that AI replaced human attackers. A human operator directed the system, and conventional software and compromised access remained central. One outlet's URL says "fully automate," but its body text says "automate" with limited human effort.

• Attribution. The link to cartels is presumed rather than established, and Straiker gave no evidence for it.

• Independent verification. Victims are anonymized and sensitive material is redacted, so findings cannot be checked independently from the published report.

How the AI Companies Responded

Per Straiker and the coverage, Anthropic revoked the compromised credential used to reach Claude after being notified, which cut off the swarm's access to its models. xAI was notified and is investigating activity tied to the Grok miners. Straiker said it was still trying to reach Moonshot AI, and none of the articles reviewed include a Moonshot statement. Straiker said both providers responded quickly.

The case shows a limit of model-level safeguards: blocking one provider does not stop an operation that can switch models, use local code or rely on conventional tools.

A Related Case in South Korea

Separately, one outlet noted that CrowdStrike described attacks on South Korean financial institutions in October 2026 using an AI tool called ARTEX, with local media putting the exposed customer data at about 68,000 people. That figure is a media report, not a confirmed official count, and it is a different campaign.

Why This Matters for Crypto Users

You may not bank in Latin America, but the pattern applies to anyone holding value online:

• Credentials are the prize. The reported inventory was passwords, API keys and access tokens. Crypto accounts and exchange API keys are the same kind of target.

• Scale favors attackers. AI lets a small team probe far more targets, so a low-profile account is not automatically safe.

• Phishing and fake tools get better. AI can produce convincing messages and fake sites. Bitrue's guide on how to avoid fake crypto tokens with contract address checks covers one common trap, and the habit of verifying before you click or sign applies everywhere.

Practical Steps You Can Take

1. Use two-factor authentication with an authenticator app or hardware key, not SMS where possible.

2. Use unique, long passwords in a password manager.

3. Limit API keys. Create keys with the least permissions needed, disable withdrawal rights unless essential, restrict by IP where the platform allows, and delete keys you no longer use.

4. Enable anti-phishing and withdrawal protections that your exchange offers, and confirm addresses with whitelists.

5. Verify links and apps. Use official URLs, not search ads or messages, and never share a seed phrase or login code with "support."

6. Review account activity regularly and act on alerts immediately.

7. Start small with any new tool, and understand what it can access.

For organizations, Straiker and secondary summaries recommend rotating credentials, enforcing least privilege, monitoring identity and payment systems, keeping an inventory of AI agents and the tools they can reach, and testing whether agents can be misused.

Where Bitrue AI Fits In

AI is dual-use: the same kind of models that can be misused for attacks are also used to help people analyze markets. Bitrue AI is one example on the trading side, and it is not a security product and has no connection to this incident. According to Bitrue's pages, it is a no-code strategy tool that analyzes market conditions, generates strategies with entry, take-profit and stop-loss levels, and explains its reasoning before you decide.

 Bitrue's materials say it requires no coding, API keys or third-party software, which means one fewer API key to create and protect, although that does not replace protecting your account with strong passwords and two-factor authentication. You review the strategy, set the amount and can stop it at any time, and Bitrue says it does not guarantee profits.

To see how it works, open Bitrue AI and the Bitrue AI Strategy page, and read Bitrue's guide to using Bitrue AI and its explanation of whether Bitrue AI is a trading bot. 

If you do not have an account yet, you can register here. Current promotions are listed on the Bitrue AI campaign page and the trading campaign page. Check the terms, and do not trade just to qualify for a reward.

Conclusion

The Straiker report is a credible warning, but it is a single vendor's account of a group's own files, with victims unnamed and the number of actual breaches unknown. What it shows is real enough: attackers can orchestrate several commercial models, switch when one refuses, and scale a small team across hundreds of targets. For crypto users, the response is basic and effective: protect credentials, limit API access, verify everything and treat any tool, AI-powered or not, with the minimum access it needs. This is not financial advice.

FAQ

Did hackers really use Claude and Grok to attack banks?

Straiker says a group it tracks as AiPT-67, linked to CyberCartel, used Claude, Grok and Kimi in a framework targeting more than 400 Latin American financial organizations. The findings come from Straiker's access to the group's files and have not been independently verified.

Were 400 banks hacked?

Not confirmed. The figure refers to targets, which include banks, payment providers and fintechs. The number of successful breaches and total losses have not been disclosed.

What did Anthropic and xAI do?

Per Straiker, Anthropic revoked the compromised credential used to access Claude, and xAI was notified and is investigating activity tied to the Grok miners.

Was the attack fully automated by AI?

Reports say no. Humans, scripts and conventional tools remained central, and several write-ups note the case does not show AI replacing human attackers.

How can I protect my crypto accounts?

Use authenticator-based two-factor authentication, unique passwords, least-privilege API keys with withdrawals disabled where possible, anti-phishing settings, and verify links before clicking.

Disclaimer: The views expressed belong exclusively to the author and do not reflect the views of this platform. This platform and its affiliates disclaim any responsibility for the accuracy or suitability of the information provided. It is for informational purposes only and not intended as financial or investment advice.