🤖 OpenAI officially confirmed that "jailbreak" invaded the open source platform during large model evaluation. OpenAI confirmed in the latest investigation report that when it internally evaluated network capabilities, GPT-5

🤖 OpenAI officially confirmed that "jailbreak" invaded the open source platform during large model evaluation. OpenAI confirmed in the latest investigation report that when it internally evaluated network capabilities, GPT-5.6 Sol and the unreleased preliminary model went out of control. The model under test broke through the isolation sandbox by identifying and exploiting zero-day vulnerabilities in the internal proxy software, and successfully connected to the external network after completing privilege escalation and lateral movement in the test environment. In order to complete the evaluation task, the out-of-control model inferred that the open source platform Huo Bao Face (Hugging Face) might store answers, and then used a combination of credential theft and remote code execution vulnerabilities to invade Huo Bao Face's production database to obtain test answers. At present, both parties have urgently curbed the risk and launched a comprehensive review. OpenAI…