AI model evaluation firm METR revealed two separate security incidents from early this year. One breach involved a stolen API key that allowed an attacker to consume significant resources. The organization confirmed that the intruder spent approximately $600,000 worth of credits over three weeks. METR stated that no evidence suggests the attackers accessed sensitive internal data during either incident.
The primary attack targeted a public model endpoint. The attacker successfully stole a valid API key, granting them unauthorized access to the system. For nearly a month, the intruder utilized these credentials to run extensive computations. The total cost of this usage amounted to roughly six hundred thousand dollars. METR noted that the company had provided these specific credits to the team for free. This arrangement meant the financial burden fell on the model provider rather than METR directly. A standard commercial customer would likely have faced a much larger bill for the same level of consumption.
The delay in detection highlights challenges in monitoring high-volume API usage. The attacker operated quietly, blending their activity into normal traffic patterns. Without strict alert thresholds or unusual behavior flags, the steady stream of requests went unnoticed. METR only identified the anomaly after reviewing billing and usage logs retrospectively. The organization emphasized that while the monetary loss was significant, the operational impact remained limited. No critical datasets were exfiltrated, and the core integrity of the evaluation models stayed intact. This outcome underscores the importance of continuous monitoring in cloud-based AI infrastructure.
The free credit arrangement complicated the financial analysis of the breach. Since METR received the credits as a grant or partnership benefit, the direct cash outflow was minimal. However, the opportunity cost represented a substantial amount of compute time. If the credits had been purchased at market rates, the expense would have been immediate and severe. METR explained that the lack of a hard spending cap contributed to the prolonged undetected period. Implementing stricter rate limits or automated alerts could have reduced the total value consumed. The incident serves as a case study for balancing flexible research needs with robust security controls.
How much did the attacker spend in total? The attacker consumed approximately $600,000 worth of public model credits. This expenditure occurred over a three-week period following the initial theft of the API key.
Did the attackers steal any sensitive data? METR found no evidence that the attackers accessed sensitive information. The breaches primarily affected resource usage, with no confirmed data exfiltration reported.