Cyber Crime at Scale: Report Details How Large Language Models Aid Criminals 

Anthropic, the company behind the large language model (LLM) Claude, released a report on Wednesday detailing 10 case studies of how people have used the artificial intelligence (AI) platform with criminal intent.

In each case, once discovered, Anthropic took steps to cut off access and, when appropriate and available, alert authorities of the activity.

The report noted the case studies are derived from uses of Claude specifically, but other LLMs are likely leveraged in similar ways. The report listed these broad areas of nefarious LLM use:

• Agentic AI systems are being weaponized. AI models are themselves being used to perform sophisticated cyberattacks—not just advising on how to carry them out.
  
  
• AI lowers the barriers to sophisticated cybercrime. Actors with few technical skills have used AI to conduct complex operations, like developing ransomware, that would previously have required years of training.
  
  
• Cybercriminals are embedding AI throughout their operations. This includes victim profiling, automated service delivery, and in operations that affect tens of thousands of users.
  
  
• AI is being used for all stages of fraud operations. Fraudsters use AI for tasks like analyzing stolen data, stealing credit card information, and creating false identities.

Each of the case studies provides details on how criminals are using LLMs to attack companies, governments, and individuals. The following are short summaries of three of the case studies with particular interest for corporate security professionals.

Victim Profiling

A person on a Russian-speaking forum leveraged Claude and Anthropic’s Model Context Protocol, a tool that acts as an interface between databases of information allowing LLMs to access and use the information, to find and build profiles on potential malware or hacking targets.

The user was able to access and analyze browser usage patterns of individuals to identify potential security vulnerabilities or possible areas of exploitation based on their browsing history. It built complex profiles of individuals that included ranked lists of user interests, as well as behavioral patterns.

“This operation demonstrates how AI transforms stolen data analysis, moving beyond simple extraction to behavioral profiling and victim prioritization,” the report said.

Stealing Credit Card Credentials

A Spanish-speaking person used Claude Code, the code analysis part of Anthropic’s LLM, to build and maintain an invite-only website designed to steal people’s credit card information for the purpose of reselling it.

The user built an application that rotated between multiple credit card validation services, included failover mechanisms, and had other sophisticated means to fool antifraud measures from card-issuing institutions. On the backend, the thief designed a method of batch processing and sale of card data that included strategic delays to disguise what might be seen as inadvertent activity rather than malicious activity.

“The multiservice validation and evasion techniques represent a concerning evolution in carding operations, potentially increasing the scale and effectiveness of credit card fraud,” the reported noted.

Vibe Hacking to Extort

Vibe hacking is a term derived from the LLM-inspired vibe coding, which gave people with only rudimentary knowledge of coding the power to create sophisticated code. Vibe hacking gives people with only rudimentary knowledge of hacking the power to create sophisticated hacking schemes. In this case study, the AI advantage was less about knowledge and more about speed and scale.

The cybercriminals utilized Claude to develop sophisticated credential harvesting techniques based on reconnaissance on individuals. This allowed large-scale network penetration of at least 17 organizations in one month.

“The operation demonstrates a concerning evolution in AI-assisted cybercrime, where AI serves as both a technical consultant and active operator, enabling attacks that would be more difficult and time consuming for individual actors to execute manually,” the report stated. “The actor demonstrated unprecedented integration of artificial intelligence throughout their attack lifecycle, with Claude Code supporting reconnaissance, exploitation, lateral movement, and data exfiltration.”

Security Management has been warning of the potential dangers of easily available online data and artificial intelligence for many years.

“Combating these emerging threats requires more than conventional strategies,” Michael R. Centrella, a former U.S. Secret Service deputy assistant director, wrote in a Security Technology article earlier this year. “It demands a holistic approach that merges technological innovation, human expertise, global collaboration, and adaptive intelligence. Organizations need to develop specialized investigative units equipped with deep understanding of machine learning, generative AI, and predictive analytics.”