I suspect a large overlap between the words likely to be used in email to Stripe (support or CEO) and the words likely to be used in e.g. 419 scams. Bad Bayesian filter seeding could easily impact legitimate messages in that case.