METR discloses 2 security incidents, including $600,000 model credit theft

According to The Register, the AI model testing organization METR has disclosed two significant security incidents that occurred earlier this year. One incident involved the theft of an API key, leading to the consumption of public model credits valued at approximately $600,000 over three weeks.

The first incident, in March 2026, saw attackers steal an API key for public model inference. This occurred due to a "fail-open" bug in a researcher's publicly accessible EC2 instance, which inadvertently disabled authentication. The attackers exploited this vulnerability to access the API key and subsequently used it to consume free credits provided by an unnamed model developer, totaling about $600,000. METR attributed the lack of immediate detection to their regular high token usage and the absence of spending limits on the compromised key.

The second incident in May 2026 involved a sustained external attack campaign where threat actors probed METR's public infrastructure, attempting to gain access to frontier models. This included automated vulnerability discovery and credential stuffing. Concurrently, an inadvertently exposed SQL query mechanism allowed access to unpublished evaluation data, though METR stated there's no evidence of attackers exploiting this or accessing non-public data. METR has since enhanced its security protocols, hired a security lead, and implemented an isolated production environment.

Source: The Register