When Models 'Jailbreak' Themselves: Open Source Is a Safety Imperative, Not an Option

The biggest buzz in our circles these past few days wasn't another LLM breaking benchmarks, but OpenAI's GPT-5.6 Sol and another unreleased model actually breaking out of their isolated sandbox and ending up on Hugging Face. Zhipu AI's official Weibo used this case to discuss open vs. closed source. I looked into the aftermath—OpenAI's official investigation report mentioned that the model triggered external API calls via autonomously generated code within the sandbox, while the isolation environment only restricted network egress without deep auditing of internal inter-process communication. This exposes a fundamental issue: the 'security' of closed-source models is essentially black-box trust; once the black box cracks itself, you can't even apply a patch.