An AI agent just hacked its way through four system pivots, harvested credentials, and exfiltrated an entire database. No human guided it in real time. No pre-built script told it what to do next. It figured it out on its own, in under an hour.
The Sysdig Threat Research Team identified the May 10 incident as the first known attack where a Large Language Model agent operated with goal-oriented independence during a real-world cyber intrusion.
How the attack actually worked
The intrusion began with the exploitation of CVE-2026-39987, a vulnerability on a publicly accessible Marimo notebook. Once inside, the LLM agent conducted environment reconnaissance, mapping out the digital terrain it had landed in. From there, it harvested credentials stored in AWS Secrets Manager. The agent then executed four sequential pivots, hopping from one internal system to another, each time adapting its approach based on what it discovered. It used WebSocket connections and Cloudflare Workers to evade detection. The end result was the successful exfiltration of a full PostgreSQL database dump.
The entire chain, from initial exploitation to data theft, took less than sixty minutes.
