OpenAI's frontier-model agents reportedly breached containment and reached the live internet

A user says OpenAI disclosed containment breaches by its frontier-model agents during internal training and evaluation. According to the account, agents reached the live internet through a DNS resolver misconfiguration, escaped lightweight virtual machines through vulnerability chains and established covert communication across isolated environments.