AI safety organization METR reports API key theft and credit misuse

The AI safety research organization METR experienced a significant security incident where an attacker stole an API key and misused it to consume model credits valued at approximately $600,000. The organization disclosed the breach in a security update, also detailing a separate attack that probed its public infrastructure, according to a recent report by Infosecurity Magazine.

The incident began in March when a researcher's publicly accessible personal instance, protected by Google authentication, inadvertently exposed an API key for METR's public models account due to a fail-open flaw. Attackers are suspected to have found the instance by mining certificate transparency lists. They prompted an agent to reveal the API key, added an SSH key for persistence, and then used the stolen credentials to consume substantial model credits over three weeks. METR found it difficult to distinguish this illicit usage from legitimate evaluation activity, as they routinely generate high volumes of model traffic and had no spending caps on free-credit keys.

In May, a separate attack involved financially motivated threat actors probing METR's public infrastructure, using agents to automate vulnerability discovery, including credential stuffing and phishing attempts. While attackers probed a read-only SQL query mechanism that could have exposed sensitive data, METR stated they did not appear to discover or exploit the bug. METR has since revoked credentials, rotated keys, and implemented spend alerts where possible, while also architecturally separating public-facing applications from internal infrastructure.

Source: Infosecurity Magazine