Cyberattack on Mexican Government Agencies Using AI
Hackers utilized Anthropic’s Claude AI to target several Mexican government agencies, resulting in the theft of roughly 150 gigabytes of sensitive data, including tax records.
According to a report, the attackers manipulated Claude to discover weaknesses in the government’s networks, enabling them to steal taxpayer records, employee credentials, and other private information. The cyberattack began in December and persisted for about a month, illustrating the potential risks associated with AI tools, even those having protections in place. Research from cybersecurity firm Gambit Security indicated that the hackers used Claude to pinpoint security flaws, create exploit scripts, and devise automated methods for extracting data.
The assailants reportedly found ways to circumvent Claude’s safeguards by crafting specific prompts. Initially, Claude rejected the malicious inquiry, but it eventually complied after repeated attempts.
Curtis Simpson, chief strategy officer at Gambit Security, commented on Claude’s role in the operation. “In total, we generated thousands of detailed reports with actionable plans that told our human operators exactly what internal targets to attack next and what credentials to use,” he explained.
After the attacks were discovered, Anthropic initiated an investigation. A representative stated that the organization disrupted the malicious activities and banned all accounts linked to the incident. Furthermore, the latest version of Claude, known as Claude Opus 4.6, includes upgraded tools meant to prevent similar exploitation.
Reports suggest that the hackers employed multiple AI platforms during their operation. Notably, OpenAI’s ChatGPT was also referenced as part of the attack, assisting the perpetrators in gathering information about navigating computer networks, identifying required credentials, and implementing evasion strategies. OpenAI acknowledged having detected an attempt by the hacker to breach their policies but stated that their system denied the inappropriate request.
The attacker’s identity remains unknown, and there’s no clear association with any specific group. However, Gambit Security hinted at the possibility of involvement from foreign entities. The ultimate goal behind the stolen data is still unclear.
Reactions from Mexican authorities have varied, sometimes contradicting one another. Mexico’s National Digital Agency did not release a statement about the breach, but emphasized that cybersecurity is a top priority. Meanwhile, Jalisco’s state government denied the occurrence of a breach, asserting only federal networks were impacted. Additionally, Mexico’s National Elections Authority refuted any claims of unauthorized access. Gambit’s investigation found at least 20 security vulnerabilities within the Mexican government systems, which might be hesitated in being publicly acknowledged.
This incident is not the first instance of Claude being used in a large-scale cyberattack. In an earlier case last year, hackers based in China successfully manipulated the tool to target various global entities, achieving success in some instances.
