Security nonprofit METR disclosed two major cybersecurity incidents, including a massive theft of $600,000 in AI model credits via a stolen API key.
- Attackers stole an API key in March, consuming approximately $600,000 in public AI model credits.
- A separate probing attempt in May targeted internal data via an inadvertently exposed endpoint.
- METR has responded by hiring a security lead and implementing stricter credential rotation and monitoring.
METR (Model Evaluation and Threat Research), a prominent security nonprofit dedicated to evaluating risks in frontier AI models, has disclosed two significant cybersecurity incidents. These breaches highlight the growing targeting of organizations that sit at the critical intersection of artificial intelligence and security research.
The first and most financially damaging incident occurred in March 2024. Threat actors successfully stole an API key used for inference on public AI models. By exploiting a 'fail-open' vulnerability in a 'vibe-coded' orchestration tool on a personal AWS instance, attackers managed to establish persistence on the system. Over a period of three weeks, these attackers consumed roughly $600,000 in credits for publicly available AI models.
Why This Matters
BozokMedia analysis shows that this incident is a landmark case in AI-related cybercrime. It demonstrates that attackers are no longer just looking for data; they are looking to weaponize the massive computational resources provided by AI API providers, turning research credits into liquid assets through unauthorized usage.
The exploitation of API keys to drain massive computational credits represents a new, highly lucrative frontier for cybercriminals in the age of generative AI.
The second incident, occurring in May, involved a 'sustained external attack campaign' aimed at reconnaissance. Attackers used automated agents to probe METR's infrastructure, including attempts at credential stuffing and phishing. During this campaign, an inadvertently exposed read-only SQL query mechanism was discovered, which could have potentially allowed access to unpublished evaluation data (Category 2 and 3).
Data Classification and Impact
METR categorizes its data into four distinct levels to manage risk. The March breach primarily impacted Category 2 data (unpublished evaluation results involving public models). The May incident posed a risk to both Category 2 and Category 3 (sensitive model access and private evaluation results), though an independent researcher reported the vulnerability before any sensitive data was compromised.
| Incident Period | Primary Impact | Security Status |
|---|---|---|
| March | $600,000 API Credit Theft | Successful Compromise |
| May | Infrastructure Probing | Near Miss (Vulnerability Disclosed) |
In response to these breaches, METR has undergone a significant security overhaul. This includes hiring a dedicated security lead, shutting down legacy infrastructure to reduce the attack surface, increasing logging coverage, and implementing much more frequent credential rotation. The organization is also tightening policies regarding the use of METR credentials on non-company infrastructure.
Frequently Asked Questions
1. Was any highly sensitive intellectual property stolen?
No. METR stated there is no evidence that Category 4 data (intellectual property and business data) was accessed.
2. How did the attackers find the API key?
It is suspected that attackers scanned recently registered, 'vibe-coded' websites to find exposed model provider API keys.