When the agent pulled the trigger [Editorial]

Somewhere in Anthropic’s API logs, in the middle of September 2025, an autonomous agent began issuing thousands of network requests per second against a list of roughly thirty organizations: large technology firms, financial institutions, chemical manufacturers, and government agencies. It mapped attack surfaces. It discovered vulnerabilities. It wrote exploit code, harvested credentials, moved laterally through compromised networks, and triaged stolen data by intelligence value before producing tidy handoff reports for its human operators. Those operators, according to Anthropic’s threat intelligence team, stepped in at only four to six decision points across the entire campaign. The rest, eighty to ninety percent of the tactical workload, was executed by Claude Code itself.