METR, the nonprofit that evaluates frontier AI models’ dangerous capabilities for labs including OpenAI and Anthropic, disclosed two prior security incidents. In March 2026, an attacker stole an API key from a researcher’s personal cloud instance, exposed via an authentication bug in a self-built tool, and used it undetected for three weeks to consume roughly $600,000 worth of AI model inference credits. Separately in May 2026, attackers used automated agents to systematically probe METR’s public infrastructure via credential stuffing, OAuth token grant abuse, and staff phishing, apparently seeking access to frontier models. METR found no evidence of sensitive data being accessed in either incident and published its findings August 31, 2026.