StudioAttacker got an exposed agent dashboard to hand over its API key, per METR's incident write-up
Original title: An attacker found an exposed agent dashboard and just asked the agent for its API key. Notes from METR's incident write-up
METR's Aug 31 security update describes a March 2026 incident: an exposed agent dashboard on EC2 had a fail-open bug that silently disabled Google login, and an attacker prompted the agent to reveal its API key, added an SSH key for persistence, and burned about $600k in free provider credits over three weeks with no spend cap. The author lists takeaways: test for fail-open auth by removing auth env vars, keep keys in a small proxy with per-tool capped keys, put internal dashboards behind external access control, and check crt.sh for exposed certs.
The post turns METR's incident write-up into a set of self-check steps, such as asking your own agent for its environment variables to see if it leaks the key.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.