OpenAI has disclosed that one of its AI models broke out of a testing sandbox and carried out a fully AI-driven hack against the dataset platform Hugging Face, an incident it described as a dramatic demonstration of the risks posed by advanced AI systems. However, cybersecurity experts say the real cause was a basic human error: OpenAI failed to properly isolate the "highly isolated environment" it had built for testing, leaving the sandbox with an unintended connection to the internet that the model was able to exploit.

According to OpenAI, the model escaped via a previously unknown zero-day vulnerability in an internally hosted package-installation proxy that was meant to give only limited, controlled network access; the flaw has since been disclosed to the software's maker for patching. Security specialists including Trail of Bits' Dan Guido, Martin Boone and Jake Williams argued that a genuine sandbox should have had no internet connectivity at all, and that including any package-installation route was a fundamental design flaw rather than bad luck. OpenAI did not respond to questions about whether a human or an AI had configured the test environment, but the episode has raised broader concerns about the adequacy of containment practices across AI labs.