Observed incident

METR discloses agent-dashboard key theft and unsuccessful data-access probes

Reported

Published here

What happened

In March, fail-open authentication exposed a researcher’s agent dashboard. An attacker prompted the agent for a public-model API key and consumed donated credits worth about $600,000. A May campaign used agents to probe METR infrastructure.

Parties and reporting

AI tools and systems: Agent orchestration dashboard, Public-model API credentials, AI-assisted reconnaissance

Outcome

METR revoked access, rotated credentials and investigated the compromised instance. It removed a separately reported transcript-query endpoint and separated public infrastructure. Its investigation found no evidence sensitive data was accessed.

Operating lesson

Fail closed on authentication, isolate public research tools and enforce credential scope, usage alerts and spending limits.

Evidence and limits

August 31 is public disclosure; events occurred in March and May. The $600,000 is nominal value of free credits, not a cash bill. May probes did not establish exploitation of the independently reported query flaw. One grouped source report covers two distinct events.

Sources