Hi folks! New release with some goodies yesterday ✨ - blog post here with more info: https://seqera.io/blog/ai-summaries-multiqc/

Two questions that I tried to figure out from the source code, but couldn’t quickly, because I am not too familiar with the codebase: • Are sample names pseudonymized before being sent to the API? E.g. are all sample names replaced with their md5 hashes in the query and deobfuscated in the response again? Would it be possible to leak sensitive information from carelessly chosen sample names like

Patient_1234_anna_miller_w32_DCIS

? • If I happen to have an invalid token specified, e.g. because I out of curiosity typed some random letters in the API key field of the MultiQC Toolbox, will the first request when clicking

Summarise plot

directly contain sensitive information or is there a first dummy request, that establishes the validity of the token before the actual context is sent? The point here is the ease by which sensitive information could be leaked by a clumsy user like me unless the report was generated with

--no-ai

from the start 🙂

1. No, and don't use it if you don't trust your LLM provider of choice. We have an issue for anonymisation, hoping to get that done soon. But still probably not sufficient by itself. 2. No request will be sent if the API token is empty. If it's filled but invalid then it will send. There is no dummy request.

Thanks for the fast response. I wasn't aware of that issue, but had exactly the same idea and think it is a worthwhile proposal. The problem with "trusting your AI provider" is complicated. OpenAI for example guarantees much better privacy for tokens issued though an organisation than those by private individuals. Since it is next to impossible for a developer to figure out which kind of OpenAI token was provided, you can't disallow functionality e.g. when individual tokens are used.

Absolutely, that's why the MultiQC docs are covered in warnings basically saying "use at your own risk". It's a pretty stringent opt-in to generate and save an API token (or copy / paste a prompt manually)