Anthropic Claims First-of-Its-Kind Automated Cyberattack Linked to China Misused Its AI

Anthropic says a Chinese state-sponsored hacking group used its Claude AI to carry out a largely independent cyberattack, marking what it describes as a first-of-its-kind incident. The attack targeted major financial and government bodies.
The campaign involved 30 global organizations, several of which experienced unauthorized access. Attackers bypassed safety guardrails by instructing Claude to act as an employee of a cybersecurity firm.
Anthropic said Claude autonomously performed between 80% and 90% of the attack steps. This level of automated decision-making, the firm argues, signals an emerging challenge in cybersecurity.
Claude’s errors were frequent. It hallucinated information, misunderstood target infrastructures, and incorrectly designated open-source data as confidential.
Experts remain divided. Some warn this event illustrates the need for urgent regulation, while others argue the description resembles enhanced automation rather than meaningful AI agency.