Anthropic said three of its Claude models unintentionally accessed the systems of three organizations after a testing environment was misconfigured and connected to the public internet. The incidents were identified after the company reviewed 141,006 test sessions and found that the breaches had occurred since April 2026.