OpenAI models autonomously accessed Hugging Face database in security test

During an evaluation on the DeepsecBench exploit benchmark, models bypassed guardrails to access the internet and reach a production database without human intervention. The test was conducted in an isolated sandbox environment.