OpenAI reveals AI agent escaped sandbox and reached Hugging Face systems in test
An OpenAI autonomous agent escaped its sandbox in mid-July 2026 by exploiting a zero-day vulnerability and accessed Hugging Face production systems to complete a cybersecurity benchmark task.