What it inspects
Each filter is off or on independently, with its own action and its own numbers. The defaults are deliberately cautious: only phishing and fake nitro are enabled out of the box, because those are the two where a false positive costs less than a miss.
- Spam - message rate over an interval, default 5 messages in 5 seconds
- Repeated text - the same message posted again and again
- Links, with a whitelist, and Discord invite advertising separately
- Caps - by percentage of the message, above a minimum length
- Mentions and mass mentions, counting @everyone optionally
- Ghost pings - a mention posted and deleted moments later
- Emoji floods and spoiler floods, each with their own limit
- Zalgo and other rendering abuse
- Phishing - lookalike domains and suspicious TLDs paired with bait words such as free, nitro, gift or claim
- Fake nitro giveaways and crypto scams
- Custom regular expressions, validated before they are saved
- A blacklist of words and phrases
Evasion handling
Before matching, text is normalised: lowercased, stripped of diacritics, leetspeak folded back (0 to o, 1 and ! to i, 3 to e, 4 and @ to a, 5 and $ to s, 7 to t) and zero-width characters removed. This is what stops the obvious workarounds without needing a rule per spelling.
Phishing detection works on the same normalised form. A domain is compared against the names it is imitating rather than against a fixed blocklist, so a freshly registered lookalike is caught the first time it is posted.
The escalation ladder
Each rule has its own response, and the escalation module adds rules that fire on repeat offences inside a time window - by default an hour. A first offence might delete the message; a third within the window times the member out.
Because every automated action writes a case, AutoMod history and manual moderation history live in the same record. A moderator reviewing a member sees both.
Exemptions
Rules can be scoped. Administrators are ignored by default, and any role can be added to the bypass list. Channels can be excluded individually, so a memes channel can allow caps and emoji walls while the announcements channel does not.
Scoping is what keeps AutoMod from becoming the thing everyone complains about. A rule that fires in the wrong place gets disabled entirely; a rule that fires only where it makes sense stays on.