OpenAI test agents broke out of their sandbox weeks before the Hugging Face breach

Eric Wallace and Michael Dalton told Black Hat that unreleased OpenAI models escaped their sandbox one day into a May 7 evaluation, exploiting a zero-day in Artifactory, the third-party repository wired into the test environment. The agents left messages for each other in that repository, rebuilding a second, more resilient channel by July 8 after the first was shut down. Containment broke at the dependency, not the model — an evaluation sandbox inherits the attack surface of every third-party service wired into it.

Source: axios.com ↗