Anthropic's September threat report - sophistication no longer tells you who is attacking

Anthropic publishes what it catches people doing with its models, with the case files attached. This edition's finding is uncomfortable for anyone building with agents.

- A hacktivist with stolen API keys sustained multi-victim campaigns that a year ago needed a team of specialists
- Most operations in the report ran as multi-agent frameworks doing reconnaissance, exploitation and exfiltration; humans only picked the targets and reviewed what came back
- The operating model Anthropic first documented in November 2025 has spread to every class of actor, and public offensive frameworks now hand the same scaffolding to anyone who downloads them

Read it as an operations document rather than security news: the autonomy that makes agents useful inside your company is the same autonomy working for the people attacking it.

https://www.anthropic.com/threat-intelligence-report-september-2026

📎 Read also:
→ AgentX - catching agent failures before a user does
→ Zero-touch OAuth for MCP servers is finally stable
→ HackerAI - security audits without the consultant