Anthropic has released its latest Threat Intelligence report, detailing the AI firm’s observations around how cyber-criminals are attempting to misuse and weaponise its flagship AI model, Claude.

The results are chilling. Anthropic has witnessed attempts to build large-scale extortion operations using Claude Code, the sale of AI-generated ransomware, and the corruption of agentic AI to perform cyber-attacks, not just advise on how to carry them out.  

The report warns that AI is being used to an unprecedented degree, with ‘vibe hacking’ now a real concern.

In one instance, Anthropic observed cyber-criminals committing large-scale data thefts, targeting seventeen organisations across various sectors and demanding ransom payments sometimes in excess of $500,000 (£370k) to stop the stolen data from being exposed.

That may seem like nothing new, but what sets this apart from the routine actions of hackers up until now was the extensive use of Claude Code by the attackers. 

‘Vibe hacking’ 

Anthropic said that, as well as being used to automate reconnaissance and the harvesting of victims’ credentials, the AI was allowed to make tactical and strategic decisions on the hacker’s behalf, deciding how to penetrate the system, what data to exfiltrate, and how to craft the extortion demands.

The Claude model was then used to analyse stolen financial data to determine how much a ransom should be, and help in the design of ‘visually alarming’ ransom notes that displayed on the victim’s screens.

In simulated ransom notes created after a close examination of hackers’ instructions, the Claude bot told victims, “We have gained complete compromise of your corporate infrastructure and extracted proprietary information…Do not test us. We came prepared.”

Likewise, the AI was used by another cyber-criminal to code, market and distribute several variants of ransomware, which were sold on dark web forums for as little as $400 (£300).

Without Claude, Anthropic said that the malicious actor would likely have been unable to develop functional malware, having witnessed them using it for core components, like troubleshooting encryption code, anti-analysis techniques, or Windows internals manipulation.  

This evolution in the use of general-purpose AI by threat actors marks a turning point, with similar attacks expected to become more common as agentic AI tools flood the market and lower the barrier for entry to those with less technical expertise.

Fake IT workers use Claude to scale fraud  

The report also found that Claude was used in the notorious North Korean fake IT workers fraud, helping DPRK operatives gain remote employment with Fortune 500 firms.

Here, the AI was used to craft false identities, including the creation of convincing backgrounds, as well as help with interviews and technical assessments during the application process. Anthropic said that, once hired, the North Koreans then used Claude to complete their coding and technical work. 

Recommended reading

• North Korean Threat Actors Have Infiltrated UK Firms
• OpenAI Reveals Scale of Threat Actors Using AI to Influence Elections
  
• North Korea Stole $659M in Crypto Last Year, Says US

Again, without Claude, Anthropic claims that the fake IT workers would have been unable to maintain their façade and keep hold of the high-paying engineering roles they’d secured, which the AI firm said were being used to fund North Korea’s weapons programme.

Whereas in the past, these kinds of large-scale, state-backed operations would require months if not years to prepare and execute, LLMs like Claude have fundamentally transformed the threat landscape.  

For its part, Anthropic said it was trying to meet the threat head-on by using these examples to strengthen its safety and security measures, concluding the report by saying: “The growth of AI-enhanced fraud and cybercrime is particularly concerning to us, and we plan to prioritise further research in this area.

“We’re committed to continually improving our methods for detecting and mitigating these harmful uses of our models. We hope this report helps those in industry, government, and the wider research community strengthen their own defences against the abuse of AI systems.”