PMG AI integration

[antonin.chadima]

Hi,

this feature would be an absolute game changer.
I would like to have a functionality in PMG like the ThunderAI for Thunderbird (just the analyze for spam part).
https://github.com/micz/ThunderAI

It should pass the message to an AI LLM model and combine the result with the SpamAssassin scores.

Prompt text something like this:

Analyze the following email and determine if it is spam or not. Consider factors such as suspicious keywords, excessive promotional language, misleading subject lines, requests for personal information, and unusual sender addresses.
Provide a value from 0 (not spam) to 100 (spam) and an explanation of no more than 10 words.
In case of missing message data, set the value to 0 (not spam), and give the reason.
Generate a response in JSON format only. Do not include any additional text or explanation; provide only the JSON. Here is the format to be used:
{
"spamValue": <integer from 0 to 100>,
"explanation": "Brief explanation of your reasoning"
}
Here there are the mail information:
Sender: "{%author%}"
Subject: "{%mail_subject%}"
Html body: "{%mail_html_body%}"

You should have ability to change the weight of the AI scores.
And setup for the connection to a LLM model (OpenAI API etc...).

It should be really easy to implement.
We have a great experience with ThunderAI but not everybody uses Thunderbird...

 

[Stoiko Ivanov]

If you want to integrate something like that - you can check-out and use the custom-check-script integration:
https://pmg.proxmox.com/pmg-docs/pmg-admin-guide.html#pmgconfig_custom_check

let us know how it works out for you!

 

[CH.illig]

Thread 'KI-Mailprüfung für PMG mit Google Gemini AI'

Jul 31, 2026

Hallo zusammen,

da wir bei uns im PMG ein paar Themen mit Phishing und Spam hatten, die so durchgerutscht sind, hab ich ein kleines Custom-Check Skript gebaut, das eingehende E-Mails per Google Gemini AI analysiert.
(Ob das nun datenschutztechnisch sinnvoll ist, muss jeder selber wissen.)
Das Ganze lässt sich bequem über die PMG Web-GUI steuern.

Was das Skript macht:​

• Prüfung per Gemini AI: Analysiert Betreff, Absender und Mail-Body auf Phishing, Social Engineering & Spam.
• Steuerung via PMG GUI: Der KI-Check läuft nur für Domains/Mailadressen, die ihr in...

• CH.illig
• spam
• Replies: 2
• Forum: Mail Gateway: Installation and configuration

• 

just an output of my daily spam fight ;-)

 

[ivenae]

I wrote one, too.
It is integrated during the last check of spamassassin to prevent unneccessary scanning.

I

[TUTORIAL] Thread 'Works as a charm: Use OpenRouter AI Service to get spamscore'

Sep 6, 2026

As described in some articles here, I’ve put together a few custom solutions to get rid of as much spam as possible as early as possible — including a Postfix policy service (Policyguard).

The small number of emails (<< 5%) that ultimately end up with a SpamAssassin score between 2 and 7 are the ones I wanted to double-check using AI.

A good fit here is the currently very strong yet inexpensive DeepSeek Flash v4 via OpenRouter. Without “thinking,” it operates quite fast (almost always < 10 seconds) and is extremely cost-efficient for evaluating emails.


How does the service work?...

• ivenae
• Replies: 1
• Forum: Mail Gateway: Installation and configuration

 

[george le]

Hi,

this feature would be an absolute game changer.
I would like to have a functionality in PMG like the ThunderAI for Thunderbird (just the analyze for spam part).
https://github.com/micz/ThunderAI

It should pass the message to an AI LLM model and combine the result with the SpamAssassin scores.

Prompt text something like this:

Analyze the following email and determine if it is spam or not. Consider factors such as suspicious keywords, excessive promotional language, misleading subject lines, requests for personal information, and unusual sender addresses.
Provide a value from 0 (not spam) to 100 (spam) and an explanation of no more than 10 words.
In case of missing message data, set the value to 0 (not spam), and give the reason.
Generate a response in JSON format only. Do not include any additional text or explanation; provide only the JSON. Here is the format to be used:
{
"spamValue": <integer from 0 to 100>,
"explanation": "Brief explanation of your reasoning"
}
Here there are the mail information:
Sender: "{%author%}"
Subject: "{%mail_subject%}"
Html body: "{%mail_html_body%}"

You should have ability to change the weight of the AI scores.
And setup for the connection to a LLM model (OpenAI API etc...).

It should be really easy to implement.
We have a great experience with ThunderAI but not everybody uses Thunderbird...

This is an interesting feature request, but integrating cloud LLMs (like OpenAI) directly into a core mail gateway pipeline like PMG introduces two major hurdles:

1. Data Privacy & Compliance: Sending inbound/outbound corporate email bodies to a 3rd-party API creates huge GDPR/privacy compliance issues for many enterprises.
2. Latency & Cost: LLM API round-trips can take several seconds per email, which can bottleneck mail queues during high-volume spikes.

However, an ideal architecture for this would be a custom SpamAssassin plugin running a lightweight, self-hosted local model (e.g., via Ollama/vLLM) to keep latency low and data on-prem. That way, PMG can simply assign a weighted SA score based on the local model's output without modifying PMG core code.

 

[ivenae]

My solution support local ollama models. You just need a graphic card to run an LLM with enough deciding power.

My solution supports only sending unclassified "on the fence" mails to AI – this is a core design feature, because AI is – as you said – one of the most time and operational cost consuming scanner/filter.

I understand the data protection concern, but i also understand, why people do not want to deliver spams, but classify those mails beforehand.

You can just send the subject, and if you send the body, you cut the first 3 - 5 lines of text instead of send the full body. This combined with only checking 5 % of your mail volume, the "on the fence" mails is a fair trade off.