Security Nonprofit Loses $600,000 in AI Credits After API Key Theft

A credential breach at the AI Model Evaluator (METR) platform has resulted in the unauthorized consumption of roughly $600,000 worth of public AI model credits belonging to a security‑focused nonprofit organization.

According to the investigation, threat actors obtained a valid API key used by the nonprofit to access METR's suite of language‑model evaluation tools. With that key, the attackers were able to run thousands of model queries, quickly exhausting the prepaid credit balance that the organization had allocated for research and development purposes.

METR, an independent initiative that provides open‑source tools for testing the robustness and safety of generative AI systems, has become a popular resource for cybersecurity teams seeking to assess emerging threats. The nonprofit that fell victim to the theft relies on METR to evaluate how AI could be leveraged in malicious campaigns, making the loss both a financial and operational setback.

The breach came to light when the nonprofit's finance team noticed an unexpected spike in credit usage on their METR dashboard. An internal audit confirmed that the activity did not correspond to any authorized projects, prompting an immediate revocation of the compromised key and a coordinated inquiry with METR’s operators.

Incidents of credential theft targeting AI credit accounts are on the rise, as cybercriminals recognize the monetary value of large language model compute resources. Similar attacks have been reported against cloud‑based AI services, where stolen keys enable attackers to generate content, scrape data, or launch phishing campaigns at scale without bearing the cost themselves.

The episode underscores the growing need for robust key management practices among organizations that depend on AI platforms. Experts recommend rotating API secrets regularly, employing hardware security modules, and monitoring usage patterns for anomalies that could indicate abuse.

METR’s team has pledged to work with the affected nonprofit to assess the full impact and explore possible reimbursement avenues. Both parties are also cooperating with law‑enforcement agencies to trace the actors behind the theft, though attribution in such cases often remains challenging.

As AI services become more embedded in security research, the incident serves as a cautionary tale for the sector: safeguarding access credentials is now as critical as protecting the data those tools analyze. Ongoing discussions within the cybersecurity community are likely to focus on establishing industry‑wide standards for credential hygiene and incident response to prevent future financial losses of this magnitude.