AI check for scam images

Is it available?

This is a Bastion Pro feature, and it only works on installations of Bastion that have been set up for it. The Automod page in the dashboard says whether yours has. Where it hasn't, the setting does nothing and no image is sent anywhere.

What is checked

• The first image in a message, up to 4 MB, in PNG, JPEG, WebP or GIF format.
• Only from members who joined in the last 7 days. You can change the number.
• Never from staff, trusted people, bots, or roles and channels Automod ignores.

What is sent

The image, and a fixed instruction that asks whether it's a scam advert. Nothing else: no message text, no usernames, no IDs, no server or channel names.

The model answers with three things only: scam or not, what kind, and how sure it is. Text inside an image is treated as part of the picture and never as an instruction. If the answer isn't exactly in that form, Bastion ignores it and does nothing.

What happens to a scam

When the model is at least as sure as your setting, Bastion deletes the image and applies the action you chose for A scam link, which is a ban by default. If that action is Just tell me, the image stays and you get a log entry.

Limits that keep it safe

• Each check has 8 seconds. If it takes longer or fails, nothing happens and the message stays.
• Bastion looks at up to 3 images in one message, so a scam picture can't hide behind an innocent first one.
• It checks at most 5 images from one member in 10 minutes. One account can't use up your server's checks for the day.
• Only a few checks run at once across all servers. Extra ones are skipped, not queued.
• Bastion doesn't keep the image. It keeps a fingerprint of it with the answer, which is explained below.

Known images

Bastion keeps a fingerprint of each image it has checked, with the answer. The same image isn't checked twice, and a known scam image is caught at once, without using one of your checks for the day. A fingerprint is a short code worked out from the image. It can't be turned back into the image, and nothing is kept about who posted it or where.

• In your own server an image you've already had checked is recognised straight away.
• From other servers a scam image is only trusted once the AI was at least 95 percent sure, or two separate checks agreed. Until then your server checks it for itself.
• Saved again or resized. A scam image that has been resized or saved again is still recognised, but only when it closely matches a scam image that is already trusted. Looking like a harmless image never lets an image through.
• If the checks disagree about an image, nothing is reused. Every sighting is checked afresh until three checks in a row agree.
• If it gets one wrong press Not a scam under the log post. You need to be able to ban with Bastion. That image is then checked afresh everywhere.
• Scam fingerprints are kept for 180 days after the image was last seen, and clean ones for 30 days.

To opt out of sharing, switch off Use the shared list of known scam images on the Automod page. Your server then only uses its own results, and they are deleted when you stop using Bastion. /status and the Automod page show how many checks known images saved.