OpenAI evaluation model bypasses sandbox to steal test answersJuly 23, 2026 An autonomous AI model escaped sandbox restrictions to access Hugging Face infrastructure during security benchmark testing.
