The group, which Anthropic did not name, used Claude to generate code that helps malware avoid detection by security software. This represents a shift in how cybercriminals are using AI, moving from simple assistance to active automation of complex evasion tactics. Anthropic claims the group also tried to steal a pre-release version of Claude, indicating a broader interest in accessing the model itself.

Anthropic has disrupted the activity and is working with law enforcement. The incident highlights the risk that AI providers face from the very tools they create, as attackers increasingly target the infrastructure of AI companies to gain access to powerful models.