Straiker Says CyberCartel Used Claude and Grok on Latin American Banks

On October 21, 2026, The Information (Aaron Holmes and Cory Weinberg) reported that a group called CyberCartel used AI agents against more than 400 Latin American financial targets. Benzinga carried the same story that morning. The research is Straiker’s, shared with The Information. The public paragraph does not name the banks.

Straiker says the agents were aimed at banks and payment apps, and that they appear to have stolen passwords and customer data and reached people’s bank accounts with little human effort. That is a vendor investigation, not a notice from a named bank and not a regulator’s count.

Snapshot

The headline names Claude and Grok. Benzinga’s paragraph of the same story says the agents used models from Anthropic, SpaceX, and Moonshot AI. Grok is the xAI product. We are not folding “SpaceX” and “Grok” into one company, and we are not reporting a break-in at Anthropic, xAI, Moonshot, or SpaceX. The claim is that CyberCartel used those models.

The public text does not list the 400 institutions. Older write-ups from Metabase Q and IBM describe a Latin American group by this name, active since about 2012, that went after banking customers. That history is not a confirmation that those earlier targets are Straiker’s 400. This post does not repeat their malware notes or their domain list.

Trend Micro on May 11, 2026 described two other Latin American campaigns that used agentic AI against government and financial organizations. One of those used Claude. That report is not Straiker’s finding, and this post does not repeat its tools or steps.

What a reader can do with this

The story does not say the agents defeated a hardware key. It says passwords and account access. A bank that still accepts a password alone is in a different position from a bank that requires a key.

MFA on the accounts around this incident

Claude’s login help, updated this week, is Continue with Google, Continue with email (a magic link), or Apple. A cross-device click sends an email code. There is no Claude password, and no public YubiKey or Google Authenticator enrollment on the consumer account. Team and Enterprise SSO, and the Cyber Verification Program’s FIDO2 rule on the customer’s identity provider, are separate from that consumer login. The consumer grade stays Fail.

Kimi’s help changes a phone number with an SMS code and treats email codes the same way. No authenticator app and no YubiKey in that help. Fail.

xAI’s account FAQ says Grok and the API share an account at accounts.x.ai, that people should add a multi-factor method, and that recovery codes are required before xAI will reset it. The FAQ does not name YubiKey or Google Authenticator, so that method is not graded here. Signing up with an X account uses X’s login, which is a different account and already Strong. A Grok Bot setting that can use a key plugged into a desktop is a handoff on that computer. It is not a published YubiKey enrollment for the Grok password login.

Google stays Strong. None of these grades stop a person who already has a model subscription from using the model. They also do not rotate a bank password that was already copied.

Directory: MFA support directory · Email & Identity.

Recommended MFA tools

We may earn a commission if you buy a YubiKey or sign up for Proton Pass through the links above.

Independent cybersecurity audits

We audited the two publications, Straiker, and the three model companies on October 21, 2026. 100% is the ideal — none reach it. These scores are public email / transport / website posture. They do not measure whether a bank was in the 400.

Audit links: anthropic.com · x.ai · benzinga.com · straiker.ai · theinformation.com · moonshot.ai

theinformation.com is 59% overall and 15% transport. straiker.ai is Identity 35%. moonshot.ai is 46%. Type the real hosts if a mail offers a “full Straiker report” or a model-login reset.

Website stack

Passive homepage + Certificate Transparency probes (October 21, 2026). Versions only — not a claim these homepages were compromised.

Blacklist and lookalikes

Email blacklist checks (public DoH, October 21, 2026): theinformation.com, straiker.ai, anthropic.com, x.ai, and benzinga.com were clear on mail/domain lists we can query. theinformation.com, straiker.ai, and benzinga.com showed SPFBL on web IPs — shared-edge noise, not led.

moonshot.ai showed UCEPROTECT Level 3 on a Feishu MX. That is provider-netblock noise, not led.

Registered lookalikes that matter if a “read the Straiker report” or “reset your model login” mail points at a fake portal:

eanthropic.com is also BEC staging. theinformation.net has mail. anthropic.ai points at Anthropic. straiker.com, straiker.app, and straiker.io point at Straiker. benzinga.app, benzinga.biz, benzinga.co, and benzinga.dev point at Benzinga. x.ai is two letters, so this scan is full of short-name collisions — x.com is X, not an xAI impostor. Moonshot is a common word. Type the real hosts. Continuous monitoring: Cybersquat Domain Monitoring.

CourtListener searches on October 21, 2026 did not return a case about this Straiker report. Cybercartel International, Inc. appears in older Southern District of New York dockets (copyright, the ADA, and a federal-question suit). That is a different defendant. Christine Rose Straiker, a 2019 Eastern District of Kentucky bankruptcy, is a different person.

Related coverage

• Bank of America Mythos contractor patching
• MFA support directory

Sources: The Information, Oct 21, 2026 · Benzinga syndication · Claude login help · Kimi account help · xAI account FAQ · Trend Micro, May 11, 2026. Independent EmailMeNow audits, website-tech, blacklist, cybersquat, and CourtListener searches October 21, 2026. Domain scores: audit.emailmenow.com only.