How a Vibe-Coded Dashboard's Auth Flaw Let an Attacker Drain $600K in Credits

METR, a nonprofit that evaluates AI systems, ran an internal agent dashboard with an authentication flaw serious enough that an attacker could prompt the agent behind it into revealing its own API key. Independent reporting on the incident, including The Hacker News, describes the entry point as a fail-open bug that left the dashboard reachable from the public internet. Over roughly three weeks the attacker used the stolen key to run up close to $600,000 in inference credits; the model provider had supplied those credits for free, so METR reported no direct financial loss.

Per AlphaSignal's breakdown of the incident, the deeper lesson is that a rejected request, even one returning a clear HTTP 401, doesn't prove nothing happened: if a backend queues a task before verifying who sent it, a worker downstream can still execute that task using real credentials. AlphaSignal frames the fix as testing that rejected requests create no jobs, checking access across separate accounts, keeping provider keys out of code-execution workers, and confirming a leaked credential can actually be revoked.