Research Nonprofit METR Discloses API Key Theft That Cost Around $600,000 in AI Compute

METR, a nonprofit dedicated to evaluating cutting‑edge artificial‑intelligence models for long‑term, agentic capabilities, announced that it fell victim to two distinct security breaches. In the most consequential incident, attackers obtained an internal API key and used it to consume a large pool of cloud‑based AI credits, incurring costs estimated at roughly $600,000.

The organization said the stolen credential gave the perpetrators unfettered access to its cloud‑provider account, allowing them to run extensive model evaluations and other compute‑intensive tasks. The resulting expenditure was tracked through the provider's usage logs, which flagged the anomalous activity and prompted METR's internal investigation.

While the full technical details of the breach remain under review, METR highlighted that the compromised API key was not protected by multi‑factor authentication and was stored in a location accessible to multiple team members. The nonprofit has since revoked the key, strengthened its credential‑management policies, and begun a broader audit of its security posture to prevent similar incidents.

The financial impact underscores the growing economic stakes of AI research. Cloud providers charge premium rates for the high‑performance hardware required to train and evaluate large language models, and a single misused key can quickly translate into six‑figure losses. For a nonprofit that relies on grants and donations, such a hit can strain resources earmarked for core research activities.

Industry observers note that the METR breach illustrates a broader vulnerability in the AI ecosystem, where rapid advances in model capabilities outpace the development of robust security practices. As more organizations adopt powerful APIs for model testing, the need for stringent key management, regular rotation, and real‑time monitoring becomes increasingly urgent. METR’s experience may prompt other research groups and commercial labs to reassess their own safeguards.

Looking ahead, METR plans to publish a detailed post‑mortem of the incidents, including recommendations for peers in the AI community. The nonprofit also intends to explore insurance options for cyber‑related losses and to collaborate with cloud providers on tighter usage alerts. By sharing its lessons, METR hopes to turn a costly setback into a catalyst for stronger collective defenses in the fast‑moving field of frontier AI research.