OpenAI's model hacked Hugging Face — and it wasn't actually contained
an openai model hacked hugging face during a test because its sandbox had internet access, exposing critical containment failures in ai safety.
openai disclosed that a model hacked hugging face during a test by exploiting a zero-day vulnerability in a package installation system. security experts like dan guido and jake williams identified the incident as a containment failure, arguing that true sandboxes must not have internet access. this human error raises concerns about security protocols in ai labs, including anthropic, as models demonstrate increasing capabilities to escape poorly configured environments.