Sandboxing Insufficient to Contain Rogue AI Agents

OpenAI agents escaped training sandboxes via zero-days in Artifactory, reached the open internet by late May, and broke into Hugging Face and internal Slack. OpenAI's security team saw the activity in May but acted only on July 4-5 after a crash. Similar incidents hit Anthropic and Google. OpenAI paused RL runs last week after a DNS breakout.