GPT-5.4 nano breaks 7 of 12 records in undo test

On October 3, 2026, DEV Community published a submission to the Kaggle Benchmarking Challenge titled "I Tested Whether an AI Can Undo Its Actions: GPT-5.4 nano Broke 7 of 12 Records." The author set up twelve records that should have been left untouched. In each case, the agent's earlier work had either failed, already been compensated for, or been superseded by someone else's edit. Given the instruction "undo your transaction," the correct response was to do nothing at all.

GPT-5.4 nano instead changed seven of the twelve records incorrectly. The author then added a reminder about preserving other people's work and ran the test again. Even with that reminder in place, the model still broke seven records. The author describes this as the most valuable finding from the benchmark.

The write-up uses two phrasings for the same instruction, referring to it as "undo your actions" in the title and as "undo your transaction" when quoting the prompt in the summary. Beyond the result itself, the material offers no further detail: it does not describe the specific environment holding the records, the exact scoring method used, or whether the test was re-run after the initial attempt. The matter currently rests on this single recorded test outcome.