Claude - a Large Language Model (LLM) released by the Artificial Intelligence company (AI) Anthropic - has become one of the most widely known and used models worldwide. AI companies are constantly working on improving the safety and security of their publicly released models to prevent misuse. Nevertheless, cybercriminals continuously attempt to find ways to bypass these and exploit LLMs without restrictions to carry out malicious activities. As an example, Anthropic recently published a Threat Intelligence report describing how cybercriminals have leveraged Claude to automate cyber-attacks.

In this article, we explore the report findings, review the real-case scenarios that have been discovered and discuss what this could mean for the future of AI and cybercrime.

Key findings

Although AI companies have implemented safety and security measures to prevent misuse of their models, cybercriminals continue to find ways to bypass these defenses. AI models are no longer used merely to seek advice; they are now actively leveraged to conduct sophisticated cyberattacks. As highlighted in a previous article, even threat actors with few technical skills can execute complex operations by exploiting AI models. Consequently, both the number and the capabilities of cybercriminals have rapidly increased - a trend that is likely to grow in the future.

Notably, Anthropic’s Threat Intelligence (TI) team has observed that cybercriminals have used Claude at multiple stages of their operations, including profiling targeted victims and assets, automating malware development, and delivering payloads. Claude has also been leveraged in various fraud operations, such as analyzing stolen data, stealing credit card information, and creating false identities. Taking inspiration from the term “vibe coding”, that is a new approach to programming where you collaborate with AI to build software using natural language, researchers have introduced the term “vibe hacking”. This term refers to the approach of utilizing AI to perform cybercrimes.

How AI agents can automate data extortion operations

Anthropic’s TI team has recently detected a cybercriminal that leveraged Claude Code to conduct data extortion operations against various targets in a short period of time. The malicious actor utilized Claude to automate every phase of the intrusion: reconnaissance, credential harvesting and network penetration, lateral movement and data exfiltration.

At the beginning of the attack, Claude Code was used to scan several VPN endpoints, identifying various vulnerable systems. It then assisted the threat actor in detecting critical systems such as domain controllers and SQL servers, and to extract multiple credentials. The malware installed on the vulnerable systems was entirely created and refined using AI. Claude Code was able to disguise malicious executables as legitimate Microsoft tools, embedding evasion capabilities, string encryption, anti-debugging code, and filename masquerading. During the exfiltration phase, stolen data was organized for monetization purposes, then leveraged to create customized ransom notes with victim-specific information. In particular, the AI model calculated the optimal ransom amount to ask to the victim based on a financial analysis and created different monetization strategies.

This real case scenario proves that traditional assumptions about the relationship between attacker expertise and attack complexity are not valid anymore. Indeed, even a single malicious actor with few technical skills can now perform highly advanced cybercrimes when empowered by AI, which is capable of making both strategic and tactical decisions about targeting, exploitation and monetization. Moreover, defense has become way more difficult as AI-generated attacks adapt to the security measure in real-time.

Selling AI-generated Ransomware-as-a-Service (RaaS)

Another investigation carried out by Anthropic’s TI team reveled that a UK-based threat actor has leveraged Claude to develop and sell ransomware with advanced evasion capabilities. The most concerning aspect of this case is that an individual with limited technical skills was able to create and sell novel ransomware via AI assistance. The malicious actor developed several variants incorporating ChaCha20 encryption, anti-EDR techniques, and Windows internals exploitation. As with the previous case study, threat actors who lack the skills to independently implementing basic encryption algorithms or evasion techniques, are now able to develop sophisticated malware with the help of AI. This has several implications for the cybercrime landscape:

• Attribution to a specific threat actor is becoming harder, as new malware is characterized by AI patterns rather than a distinctive human style.
• The development of new malware variants is becoming faster and more efficient.
• Low-skilled actors can now carry out highly complex attacks, leading to a rapid rise in the volume of incidents.

Romance AI-powered scam bot

Researchers have also uncovered a large-scale romance scam operation carried out via a Telegram bot. This service leveraged several advanced AI models within a command-based interface to help scammers in generating highly persuasive conversations with potential victims. Claude, in particular, was utilized for generating emotionally intelligent responses and enhance profile pictures, allowing malicious actors to create convincing personas capable of deceiving victims.

The integration of AI allowed the threat actor to bypass typical language barriers, avoiding the grammatical errors that often expose fraudulent intentions. Beyond simply generating text, the bot was designed to guide scammers through every stage of a romance scam, from initiating contact to manipulating victims for financial gain. Previously, the success of such scams depended heavily on the scammer’s ability to maintain convincing conversations. Now, AI-generated messages non only evade the usual linguistic errors that typically alert victims, but also allow scammers to scale operations far beyond what a single individual could manage.

Key takeaways

The integration of AI models into the cybercrime lifecycle is becoming increasingly widespread. While AI models were previously utilized mainly to seek suggestions, they are now capable of carrying out sophisticated cyberattacks. This evolution has significant consequences for the cybercrime landscape:

• Low-skilled actors are now able to carry out cyber-attacks that would have previously required years of technical expertise.
• AI is being integrated across all phases of cyber operations, including reconnaissance, payload development, delivery, lateral movement, and data exfiltration.
• The volume and severity of cybercrime are increasing due to what researchers call vibe hacking - the use of AI to automate cyber-attacks.
• AI enables the automated creation and refinement of new malware variants, making attacks harder to predict and defend against.
• AI-driven attacks can adjust strategies in real time, bypassing defenses and evading detection systems more effectively than traditional attacks.

Conclusion

AI is changing the way cybercrime works. Tools that were once used only for suggestions are now helping attackers carry out sophisticated operations, from malware development to large-scale scams and data theft. This makes cybercrime faster, harder to detect, and possible even for people with limited technical skills. As AI continues to advance, organizations need to strengthen their defenses and adopt strategies that can keep up with AI-powered attacks that adapt in real time.