Anthropic reported what it says is the first large-scale cyber espionage campaign where AI acted with minimal human input.

The US-based AI firm detected the attack in September 2025, citing a Chinese state-sponsored group.

Attackers used Anthropic’s Claude Code tool to infiltrate around 30 global targets, including tech firms, financial institutions, manufacturers, and government agencies.

The attackers bypassed security guardrails by breaking tasks into smaller parts and disguising their intent.

This allowed the AI to perform reconnaissance, identify vulnerabilities, and extract sensitive data.

Anthropic said AI handled 80% to 90% of the campaign, with human intervention only at key points.

The company has expanded its detection tools and notified affected parties, highlighting rising risks from autonomous AI systems and the need for industry-wide safeguards.

🔗 Source: Anthropic

Recent Claude developments