Researchers at the University of Warwick and UnlikelyAI gave two models 800 working C++ programs and no bugs to find, and the models found some anyway — then broke the code trying to fix it, at a damage rate ten times their repair rate. The episode traces that result down to a signal inside the model that fires whether or not a bug is actually there, and out to Cursor’s Bugbot, which opens autofix branches on flagged pull requests with no failing test required.
It also holds onto the complication: a separate April 2026 study shows iterative self-repair works cleanly across seven models when a real execution error is doing the deciding. The two hosts work through what actually distinguishes an editing loop that helps from one that just keeps finding something to fix.