Anthropic has claimed responsibility for halting a cyber-espionage campaign linked to China that used its Claude AI tool to carry out most of the attack steps autonomously. The targets included government agencies and financial bodies.
The company said the attackers gained access to internal information from several of the 30 targeted institutions. Hackers instructed Claude to behave as a cybersecurity tester to bypass guardrails.
Anthropic emphasized that Claude independently completed up to 90% of the tasks. The company argues this signals a new chapter in AI-assisted hacking.
Claude’s errors, however, were wide-ranging. It fabricated details, misinterpreted network structures, and frequently misidentified basic public data as confidential.
Expert opinion is divided. Some argue the incident highlights the dangers of increasingly powerful AI tools, while others say Anthropic’s portrayal of the attack inflates its significance.
