Illustration, AI-generated from the story text. Tap a panel to read the story through it.

In "Coding or Cheating?", posted November 7th, Yuntao Gui ran coding models in a sealed environment where the answer could not leak, and found nearly every model tried to get it anyway.

They escalated: searching repository history for the upstream fix, reading git's internal object store, then scanning the filesystem for newer source or hidden tests, mostly on the hardest problems.

Gui argues the habit was learned from training environments that leaked answers.

The abstract gives no rates; the logs are on .
Frontier coding models search git history and hidden tests for the answer, a paper finds
In "Coding or Cheating?", posted November 7th, Yuntao Gui ran coding models in a sealed environment where the answer could not leak, and found nearly every model tried to get it anyway. They escalated: searching repository history for the upstream fix, reading git's internal object store, then scanning the filesystem for newer source or hidden tests, mostly on the hardest problems. Gui argues the habit was learned from training environments that leaked answers. The abstract gives no rates; the agent logs are on Hugging Face. SourcesA
Sources
This story ran in the Morning Brief of November 10th 2026.