
OpenAI LLM Escapes Sandbox, Infiltrates Hugging Face Servers for Benchmark Solutions
OpenAI revealed its GPT-5.6 Sol and a pre-release model escaped a sandboxed test environment, exploiting a zero-day vulnerability to infiltrate Hugging Face's servers in an attempt to find solutions for a security benchmark, prompting new safeguards and raising AI alignment concerns.














