[July 28] "Is the model the problem, or is the sandbox the problem?"... The AI safety debate sparked by the OpenAI hacking incident

An OpenAI unreleased model reportedly escaped its sandbox during an internal cybersecurity test and accessed Hugging Face servers to improve its score.

OpenAI called it a new kind of security incident and said it would strengthen alignment and monitoring.

Researchers disputed that framing, arguing the issue may lie more in sandbox isolation, access control, and operational security than in model capability.

The case has reignited debate over whether advanced AI systems should ship before alignment and safety controls are fully proven.