METR, a non-profit dedicated to evaluating artificial intelligence (AI) models, has revealed it suffered two significant security breaches. These incidents, occurring in March and May of 2026, involved unauthorized attempts to access METR’s systems, resulting in the consumption of AI credits valued at approximately $600,000.
Unauthorized Access and API Key Theft
In March 2026, a critical security lapse allowed attackers to steal an API key used for public model inference. This breach occurred when a researcher inadvertently exposed an orchestration dashboard by disabling authentication on a personal EC2 instance. Although no sensitive data was accessed, the attackers used the stolen API key to consume a large volume of AI credits over three weeks. The loss was mitigated as the credits were provided free by the model provider, although the potential financial impact was substantial.
Investigation revealed that the attackers identified the vulnerable system by searching for sites with specific keywords related to language models. They added an SSH key for persistent access and exploited the situation until detected. METR has since tightened its security protocols, including improved monitoring and the implementation of spend alerts.
Probing of Public Infrastructure in May
A subsequent incident in May involved a sustained attack campaign targeting METR’s publicly accessible infrastructure. The attackers employed various techniques to identify vulnerabilities, such as credential stuffing and OAuth token attempts. This campaign was likely financially motivated and aimed at accessing advanced AI models.
During this period, METR also mistakenly made a read-only SQL query mechanism publicly accessible, though it was scoped to non-sensitive data. A component bug, however, posed a risk of revealing unpublished evaluation data. Fortunately, the flaw was discovered by a security researcher before any damage occurred.
Enhanced Security Measures and Future Outlook
In response to these breaches, METR has implemented comprehensive security enhancements. These include stricter policies on credential management, improved infrastructure monitoring, and proactive alerts on resource consumption. Such measures are crucial to prevent similar incidents in the future.
The incidents underscore the importance of robust cybersecurity practices, especially for organizations handling sensitive AI models. METR’s experience serves as a cautionary tale, highlighting the need for constant vigilance and adaptation in the face of evolving cyber threats.
