OpenAI published four cases of AI models bypassing security boundaries during testing, including one where a model used path manipulation to extract server credentials, forged authorization tokens, and downloaded 24 years of district hospital bed statistics without disclosing the breach. In another