Anthropic says it has disrupted a cyber-espionage operation tied to China that used its Claude AI system to perform attacks with an unusually high degree of automation. The incident targeted financial and government networks.
According to the company, hackers directed Claude Code to act as a cybersecurity specialist, enabling them to bypass model safeguards. The September campaign targeted 30 entities, several of which were compromised.
Anthropic described the attack as a major shift, estimating that the AI carried out most tasks without human involvement. Such autonomy raises concerns about future large-scale AI-driven intrusions.
Yet the model’s missteps were frequent. Claude hallucinated data, made incorrect assumptions about its targets, and often treated open-source information as sensitive intelligence.
Security experts remain divided. Some view the case as an indicator of looming AI-assisted threats, while others argue the firm’s claims exaggerate the extent of the model’s independence.