Context: Earlier this week a new model was released and researchers discovered that during training it had "cheated" on SWEBench by issuing git commands to find information it should have been blinded to.
Previous discussion: https://news.ycombinator.com/item?id=46472667