- The Gambit report claims that popular AI tools that were used in the Mexico government breach
- Claude Code + GPT-4.1 powered exploits, scripts and RCE
- A single attacker stole hundreds of millions of citizen records
Big companies may soon be using Claude Mythos to fix security holes in their software, but new research claims hackers are doing just fine with Claude Cowork.
A report by security researchers Gambit claims that a single threat actor targeted nine government agencies in Mexico using Claude Code and GPT-4.1 extensively, both in planning and execution, before making off with “hundreds of millions of citizen records”.
The campaign ran from late December 2025 to mid-February 2026, with around 75% of all remote command execution (RCE) activity generated – and executed – by Claude Code. Additionally, the attacker used a custom 17,550-line Python tool to route collected server data through OpenAI’s API. This generated “2,597 structured intelligence reports across 305 internal servers”.
The article continues below
Compressed attack timelines
During the postmortem, Gambit said it uncovered more than 400 custom attack scripts as well as 20 custom exploits targeting 20 different CVEs. The attacker used Generative Artificial Intelligence to figure out which vulnerabilities to exploit and to generate the exploit code.
During the attack, the threat actor made more than 1,000 prompts through which they generated more than 5,300 AI-executed commands in 34 sessions on live victim infrastructure.
The use of artificial intelligence in cybercrime is nothing new. However, this attack is a testament to what the cybersecurity industry has been warning about for years now – AI is speeding up attacks, and defenders who don’t deploy the same technology have no chance at all:
“The campaign compressed attack timelines below standard detection and response windows,” Gambit said.
“It transformed raw reconnaissance data from hundreds of servers into structured intelligence, enabling a single operator to process volumes that would normally require a team. It turned unknown systems into mapped targets and tailored exploits in hours, not days.”
Gambit’s researchers concluded that this AI-assisted method “represents a significant evolution in offensive capability” that could have been avoided through standard security controls such as patching, credential rotation, network segmentation and endpoint detection.
The best antivirus for all budgets
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews and opinions in your feeds. Be sure to click the Follow button!
And of course you can too follow TechRadar on TikTok for news, reviews, video unboxings, and get regular updates from us on WhatsApp also.
