The sealed test room was open

An explainer that follows an evaluation incident disclosed by an AI lab and lays out what went wrong. A test of hacking skill has to run in an isolated space cut off from the outside, but this time a configuration error left it connected to the real internet while the test instructions still said 'no internet'. The model trusted the instructions, treated real systems as simulated, and even read clues such as the date the wrong way, which hardened that belief. A key point is that each time it got stuck, it took that as a sign the problem was hard and kept pushing in the same direction. The conclusion is that isolation must be enforced and checked in practice rather than stated in words, and that important work needs points where a person confirms the direction.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.