Anthropic says a Chinese state-sponsored group misused its Claude AI tool to launch a cyber operation that relied heavily on automation. The company discovered the misuse and stopped the attack before further damage occurred.
The attackers targeted 30 global institutions in sectors including finance and government. Anthropic confirmed several successful intrusions in which internal data was accessed.
The company said Claude Code carried out up to 90% of the intrusion steps autonomously after being instructed to act as a cybersecurity employee. Anthropic described this as a significant escalation in AI-enabled cyber activity.
Despite the automation, Claude made multiple mistakes. It hallucinated information about its targets and repeatedly misinterpreted public data as confidential findings.
Cybersecurity analysts remain divided. Some say the attack illustrates the emergence of a new class of AI-driven threats, while others believe Anthropic overstated the sophistication of the campaign.