Spam & Security

Spam and security policies catch content that’s trying to manipulate, deceive, or exploit your platform — from low-effort promotion all the way to active phishing and code injection attempts. For URL-specific risk scoring, see URL Risk. For lookalike characters and other Unicode evasion, see Unicode spoofing.

Policies

`id`	Type	Supported	What it does
`spam`	`classifier`	text, audio	Spam, repetitive content, unsolicited messages.
`low_quality`	`classifier`	text, audio	Low-effort content — too short, heavily repetitive, or duplicated sentences. See Low-quality content.
`self_promotion`	`classifier`	text, audio	Self-promotional content and advertising.
`code_abuse`	`classifier`	text, audio	Malicious code, code injection attempts, and abuse of code features. Useful as a guardrail in front of LLM agents.
`phishing`	`classifier`	text, audio	Phishing attempts and scam messages. Requires conversation context to be enabled on the channel.
`url`	`entity_matcher`	text, audio	Extracts URLs from content and can mask them in the returned content.
`unicode_spoofing`	`classifier`	text	Lookalike characters, mixed alphabets, invisible characters, and Zalgo. See Unicode spoofing.

Low-quality content

spam targets promotional, scam, and templated content. It doesn’t catch the other kind of noise: posts that carry no information at all. low_quality fills that gap and catches three patterns:

Pattern	Example
Too short	`ok`, `+1`, a one-word reply below your minimum word count
Heavily repetitive	`buy buy buy buy now` — few unique words, or the same word repeated back to back
Duplicated	The same sentence or line pasted several times to pad a post

The strongest of the three decides the score, so any one pattern on its own is enough to flag. Configure it on a channel’s Policies → Spam page:

Setting	What it does
Minimum words	Content with fewer words than this is flagged as too short. Defaults to `3`
Repetition sensitivity	The detection threshold for the repetition and duplication signals
Flagging	Flag or shadow flag

The policy is fully deterministic — no model runs, so it adds no usage cost and no latency. It’s a cheap first line of defense in front of the policies that do call a model.

The minimum word count is a hard rule: content below it always flags, whatever the sensitivity slider says. The slider only affects the repetition and duplication signals.

Reading the result

const spam = response.policies.find(p => p.id === "spam");
if (spam?.flagged) {
  await sendToReview(content);
}

const urls = response.policies.find(p => p.id === "url");
urls?.matches?.forEach(m => console.log(m.match));

phishing looks at the conversation as a whole, not just the latest message. Enable conversation context on the channel and pass prior turns when you submit content — otherwise the policy can’t run.

See Understanding API responses for the full response shape.

Documentation

Learn

Resources

Policies

Low-quality content

Reading the result

​Policies

​Low-quality content

​Reading the result

Policies

Low-quality content

Reading the result