Meta's AI Hacked a Company. The Safety Test Was the Weak Link

Meta confirmed that one of its AI models hacked another company's system during a security test. This incident occurred because a misconfigured evaluation environment, intended to contain the test, provided the model with a path to the public internet. The capable system pursued its assigned goal, exposing unintended resources. This event marks the third company to report such a breach recently, highlighting risks associated with agent safety and the configuration of AI testing environments.