Second opinions, takeovers & admin actions — Discord Bot…

Nothing automatic lands unread

Before an automatic timeout, kick, ban, quarantine or role strip is applied, the messages that caused it are read back. The rules counted; this reads. Six fast messages about a goal and six spam messages look the same to a counter, and the member on the wrong end of that has no recourse until staff wake up.

When the reading positively disagrees, the action is held and staff get an alert saying what it looked like instead. Silence, uncertainty or an outage all leave the original decision alone — this can never make the bot softer by failing.

A moderator who typed the command is never second-guessed. Only automatic actions are reviewed.

The trusted account that is not

The realistic attack on an established server is not thirty fresh alts; it is one long-standing member whose token was stolen, now pitching a scam to people who trust them. Every security bot exempts that account by definition. Sweet keeps a sample of how established members normally write here, and when a link-carrying message stops sounding like its owner and has the shape of a stolen-account pitch, staff get a critical alert and the message carries heat. It never bans on this alone.

Administrative actions, judged in context

Anti-nuke counts destructive actions and reacts once enough have happened, which means watching four more channels disappear before the rule agrees with what is already obvious. Each action is also judged on its own: who did it, how long they have been trusted, what was granted. When a single action reads as a takeover rather than housekeeping, anti-nuke treats it as its threshold immediately.

Whitelisted staff are still judged, but never punished — being trusted exempts you from the action, not from being looked at. When a trusted account does something that reads as a takeover, staff are told and nothing else happens.

Coordinated posting

Five accounts posting one message each never trip a per-user threshold, which is exactly why advertising rings post that way. When several accounts are flagged close together, their messages are compared as a set; if they are running one campaign, every participant carries the weight of the campaign rather than of their single message.

Names

Word lists read messages and ignore the two places a name sits permanently in front of everyone. Nicknames and voice room names are read as well, and a name that is a slur, sexual, impersonating staff or carrying an invite is reset.