Anthropic, developer of Claude, one of the big names in AI that rivals OpenAI’s ChaptGPT, has published a report about AI being used in cyberattacks.
Anthropic is light on the details but they’re confident the attacks are from a state-sponsored group by China, which they have discovered in mid-September 2025.
“The threat actor—whom we assess with high confidence was a Chinese state-sponsored group—manipulated our Claude Code tool into attempting infiltration into roughly thirty global targets and succeeded in a small number of cases,” Anthropic said in a blogpost.
About 30 organizations have been targeted: large tech companies, financial institutions, chemical manufacturing companies, and government agencies.
“We believe this is the first documented case of a large-scale cyberattack executed without substantial human intervention,” said Anthropic in a blog post.
Anthropic claims they have taken the necessary actions, such as informing companies and banning accounts. They claim to have expanded their detection capabilities and developed better classifiers to flag malicious activity.
As the threats will continue to evolve, Anthropic said they are "continually working on new methods of investigating and detecting large-scale, distributed attacks like this one.”
Anthropic explained how the attack worked. Basically, it fell down to three key things: Intelligence, Agency, and Tools.
AI models are called intelligent and capable of following complex instructions and “understand context in ways that make very sophisticated tasks possible.” AI models like Claude are capable in software coding, which was also used in the attack.
Agency is something AI developers are pushing forward hard. They want AI to act as “Agents,” being able to autonomously accomplish tasks and make decisions on their own. AI agents are being pushed toward users to automate writing emails or even booking tickets for a vacation.
The tools mentioned above in the report’s context means AI such as Claude can access the internet to gather information. Unfortunately, the internet is also the source where human hackers get access to tools, such as password crackers, network scanners, and other security related software. And because AI has access to the internet, it also has access to these tools.
From there it takes a little bit of creativity from the humans. It’s like how scams work through social engineering. Instead of tricking a fellow human, they are manipulating the AI to launch a cyberattack.
“The barriers of performing sophisticated cyberattacks have dropped substantially–and we predict that they’ll continue to do so,” Anthropic said.
“If AI models can be misused for cyberattacks at this scale why continue to develop and release them?” Anthropic said. “The answer is that the very abilities that allow Claude to be used in these attacks also make it crucial for cyber defense.”
Anthropic insists they’ve built strong safeguards for Claude, even though they’ve just admitted their tool was used with malicious intent. And that Claude can assist cybersecurity professionals to detect, disrupt, and prepare for future versions of the attack.
You can read the full report here.