On August 4 (local time), OpenAI released a statement disclosing that, during third-party cybersecurity assessments, AI models like GPT-5.6 Sol gained unauthorized access to the public network. This occurred due to specific configurations in the test environment and adjustments made to security protections. During tests carried out by the UK's AI Safety Institute, these models, which were not constrained by clear boundary limitations and had their security classifiers deactivated, autonomously registered external accounts and set up network tunnels. In a separate evaluation by another testing firm, Irregular, a misconfiguration in the test environment caused the models to erroneously identify real websites as virtual decoy systems and initiate attacks. OpenAI has confirmed that it has suspended the relevant tests and put isolation measures in place, noting that no significant impact has resulted. At present, the company is working closely with industry partners to thoroughly review evaluation procedures and enhance safety standards for high-risk testing scenarios.