1-minute read
What a defender needs to know
An internet-facing AI-agent dashboard exposed a METR model-provider API key. The application had a fail-open flaw that silently disabled Google authentication. An attacker prompted the agent to reveal the key, added an SSH key to the personal EC2 instance, and used the credential for three weeks to consume model credits worth approximately $600,000. The provider had granted those credits to METR for free.
METR says the key reached public models and that internal and consultant-assisted forensics found no compromise beyond the single credential. High-volume evaluation traffic, incomplete visibility into rate-limited requests and the absence of a provider spending limit delayed recognition.
Inventory public agent dashboards, personal cloud workloads and experimental services that carry organizational credentials. Record each service’s authentication behavior, owner, reachable routes and key inventory. Alert on new SSH keys, authentication changes, sustained rate-limit activity, unusual provider usage and requests from unapproved environments.
Preserve provider usage and rate-limit logs.
If exposure is suspected, image the host, revoke access, rotate every credential present and reconcile host activity with provider usage. The verification target is a fail-closed denial on every unauthenticated route, no organizational credentials on unapproved infrastructure, an owner for every active key and alerts that surface abnormal use before it persists.