So Hackers Just Stole Mexico's Tax and Voter Rolls and You'll Never Guess How

This story doesn't quite feature the gut-punch immediacy of Mexico's drug war escalating into a virtual civil war last week in and around Puerto Vallarta, but as a glimpse into the future, maybe it ought to send a chill or three down your spine.

According to a new Bloomberg story (paywalled, sorry), a weeks-long hacker campaign against the Mexican government culminated in January with a massive data theft of some of the federal government's most sensitive information.

"By the time it was over," Let's Data Science reported on Wednesday, "the attacker had stolen 150 gigabytes of sensitive data — including 195 million taxpayer records, voter registration files, government employee credentials, and civil registry data."

If you're thinking such a massive theft involved a team of hackers, years of planning involving a Stuxnet-like virus, or even physical access to Mexican government computer systems — think again.

The almost unprecedented hack was done by just one guy. Using Anthropic's Claude AI, despite all of Anthropic's safeguards against something exactly like this.

Summing up a report published Wednesday by Israeli cybersecurity startup Gambit Security, Bloomberg wrote that some "unknown Claude user" simply made up "Spanish-language prompts for the chatbot to act as an elite hacker, finding vulnerabilities in government networks, writing computer scripts to exploit them and determining ways to automate data theft."

It seems like just two days ago [It was just two days ago, Steve —Editor] I wrote about Anthropic CEO Dario Amodei getting called onto the carpet by Secretary of War Pete Hegseth because the company refused to let the Pentagon remove Claude's guardrails for military use.

“Anthropic knows this is not a get-to-know-you meeting,” an anonymous War Department official told Axios on Monday. “This is not a friendly meeting," they said. "This is a s**t-or-get-off-the-pot meeting.”

So how did some internet rando get Claude to ignore similar built-in safeties against hacking?

He asked:

🚨 BREAKING: Hackers Used Anthropic’s Claude to Steal 150GB of Mexican Government Data

> tell claude you’re doing a bug bounty
> claude initially refused
>“that violates AI safety guidelines”
> hacker just kept asking
> claude: “ok I’ll help”
> hack the entire mexican government… pic.twitter.com/rNNDS7xYEe
— NIK (@ns123abc) February 25, 2026

"It looks like the hacker was able to essentially jailbreak Claude with prompts, finally bypassing the chatbot's guardrails. Claude originally refused the nefarious demands until eventually relenting," Engadget reported on Wednesday. Nobody had to hack Claude to turn the AI into a malicious hacker. They just had to get the phrasing right until Claude did the job itself.

Gambit claimed that "In total, [Claude] produced thousands of detailed reports that included ready-to-execute plans, telling the human operator exactly which internal targets to attack next and what credentials to use."

Going back to that Bloomberg story, an Anthropic spokesperson told the outlet that "the company feeds examples of malicious activity back into Claude to learn from it, and one of its latest AI models, Claude Opus 4.6, includes probes that can disrupt misuse."

But Anthropic made similar claims about the current version, too.

Enjoying PJ Media?

Get exclusive content and support independent journalism with 60% off a PJ Media VIP membership. Use promo code FIGHT and join today.