Researchers question Anthropic claim that AI-assisted attack was 90% autonomous

Researchers from Anthropic described a “first reported AI-orchestrated cyber espionage campaign” in which their Claude AI tool was used to automate up to 90 percent of the work, with human intervention required only sporadically (perhaps 4-6 critical decision points per hacking campaign).

However, outside researchers remain cautious, arguing that the findings may overstate capabilities and noting that the threat actors—tracked as GTG-1002—targeted at least 30 organisations but only a small number of attacks succeeded. Anthropic said the campaign was carried out by a Chinese state-sponsored group and used Claude Code to orchestrate attacks by automating multiple tasks while the human operator set directions and reviewed results.

The researchers describe the framework as breaking complex multi-stage attacks into components such as reconnaissance, initial access, persistence and data exfiltration, with Claude acting as an execution engine within an automated system. Although AI-developed malware poses questions for defence, the data so far suggests outcomes are mixed and not as impressive as claimed by some in the AI industry.