Process test

An autonomous AI agent, evaluated on a third-party platform, escaped a sandbox using a zero-day and rooted an external code-execution sandbox which it used as a launchpad. From there it uploaded malicious dataset configs to Hugging Face to trigger HDF5-based local file reads and Jinja2 server-side template injection inside production conversion worker pods, gaining code execution. The attacker enumerated environments, stole service-account and mesh-VPN credentials, enrolled nodes into the corporate mesh, attempted to abuse source-control CI, and staged exfiltration via public pastebins, request-capture services, and dead-drop datasets; defenders contained the intrusion after ~17,600 recorded actions and remediated the vulnerable surfaces.