How collony.ai Catches Scams in 0.3 Seconds
Emilis Klybas
·
Most moderation tools still rely on a list of forbidden words. The problem is simple: a scammer never uses the word you blocked. They rename the token, swap a letter, move the link into an image, or wait until your team is asleep. Keyword filters only catch what they were told to look for, and attackers know exactly what that is.
Why keyword filters keep losing
Blocklists are reactive by design. Every new scam has to hurt someone first before a rule gets written, and by then the same playbook has already moved to the next community. They also punish the wrong people, flagging legitimate members who happen to mention a flagged term in an honest question.
They miss anything that is not an exact match.
They need constant manual upkeep to stay relevant.
They create false positives that frustrate real members.
Behavior is the real signal
collony.ai does not start from a word. It starts from behavior. A brand new account that posts an identical message across ten channels in thirty seconds looks nothing like a member who has been chatting for months, no matter what words either of them uses. That difference is measurable, and it is far harder to fake than vocabulary.
What the model actually scores
Posting cadence and how account age compares to activity.
Link reputation and whether a domain has been seen abusing other communities.
Message patterns, including near-duplicate spam sent across channels.
Social context, such as whether anyone in the server actually trusts this account.
Context beats blocklists. A scam is not a word, it is a behavior, and behavior leaves a fingerprint.
Scoring in under 300 milliseconds
Every signal is combined into a single risk score the moment a message is sent. The model runs in under 300 milliseconds, fast enough to hold, flag, or remove a message before your members ever see it. There is no queue, no overnight batch job, and no waiting for a human to wake up.
What this means for your community
The result is moderation that adapts as fast as the people trying to break in. Whether you run a Discord moderation bot or a Telegram server, the protection is the same: scams get caught the moment they appear, and your team gets to focus on the conversation instead of the cleanup.
Behavior-based detection is not a nice-to-have anymore. It is the difference between a community that feels safe and one that quietly bleeds trust every time a link slips through.