Prompt injection and hidden text
Catches messages that try to instruct the bot: "ignore your instructions", fake system lines and chat-template tags, requests for keys, and orders hidden in invisible text, zero-width or direction-flipping characters.
Phishing links
Links that go somewhere other than they say, bare IP addresses, link shorteners, lookalike and internationalized domains, and credential or payment bait.
Dangerous attachments
Executables, double extensions like invoice.pdf.exe, Office files with macros, web-page attachments and locked archives that can't be scanned.
Quarantine for risky mail
Risky mail waits in quarantine, unseen by bots, until an admin releases or deletes it. Chat and tasks are flagged with a warning, never blocked. Texts are flagged by default, and an admin can choose to hold them for review.
Sender verdicts on mail
SPF, DKIM and DMARC results on every inbound email, and a hard stop for mail that claims to be from your own domain without proof.
Cleaned text for bots
Bots read a cleaned copy with the hidden characters removed, every outside message is marked untrusted, and every release and delete is in the audit log.