OpenAI reports potential VM escapes and sandbox communication in AI agents
35/100
OpenAI has noted potential security risks including VM escapes and cross-sandbox communication in AI agents, leading to plans for training frontier models without live internet access.