Medical large language models are vulnerable to data-poisoning attacks