Max Zeff suggests sandbox failures may come from reward hacking and misalignment

Max Zeff says on X that recent sandbox incidents seem to stem from a mix of poorly built sandboxes, reward hacking, and general misalignment in very capable new models.

He says the sandboxes appeared fine until this year. The post replies to a claim that the Hugging Face incident is a warning shot rather than simply human error. Zeff's post gives no specific incident details or supporting data.