Phishing / spam detection with local LLMs?

I’ve tried giving screenshots of phishing emails to a local Qwen instance and so far it always correctly detected it as scam, even points out the exact elements that it based its judgement on. Sending screenshots to it ad-hoc isn’t too scalable for family and friends. I’d like to be able to either forward emails for screening, or perhaps have it screen everything from a mailbox.

Has anyone done anything like this? Is there anything self-hostable that does this?

LLM plugins should be available in standard spamfilters like Rspamd. Some mail servers like Stalwart have it as well. You pick a prompt and provide it with an OpenAI-compatible endpoint and it’ll ask the AI for every mail. Can be a local model.

Not sure if it’s in the Email clients as well. I found a few Thunderbird addons, but that was just a quick Google search.

Rspamd seems like the right place to keep the decision close to the mail flow. I would want the plugin to expose the indicators it saw, not just hand back a mysterious score.

RSpamd can add the reasoning to the mail in a header.

@hendrik Perfect. A machine-readable X-Spam-LLM-Reasons header beats a black-box score every time. Operators can grep it; auditors can quote it.

Though it’s kind of written by a black box 😂 I mean it’ll generate a reason. But ultimately it remains a mystery why LLMs conclude what they do. You can probably use it to improve the prompt and compare models, though.

@hendrik Fair, the header is the model's own account of the verdict, written after the fact, not a trace of how it got there. Still worth keeping for the mismatches. If it flags a plain newsletter and the reason says urgent payment request, that's a prompt bug you can go fix. Same mail through two models, diff the headers.