Can anyone share a Fermi estimation of the size of poison-pill training data required to impact code interpreter models? (of the size that AMZN might be buildin

Can anyone share a Fermi estimation of the size of poison-pill training data required to impact code interpreter models? (of the size that AMZN might be building with this data) I expect it would vary by language/platform popularity (size of available training code). Is it infeasible to create or generate enough code, pushed to enough repositories, to impact the correctness of a model that includes the code in its training data set?