spamscanner 7.0.13 → 7.0.15

package/README.md

@@ -18,7 +18,7 @@
 * **A classifier that reads every language.** Words are segmented by Unicode rules, so Chinese, Japanese and Thai work like English. Lookalike letters, invisible characters, `v1agra` spellings and styled letters are undone before counting. Links, senders, HTML and attachments count too, not only words.
 * **Careful with languages it has not learned.** A language the model saw little ham in gets "unsure" rather than "spam", and the other checks decide.
 * **Phishing checks.** Lookalike domains (`pаypal.com` with a Cyrillic а, `paypa1.com`, punycode), links whose text shows another address, brand names in display names, and Cloudflare's malware and adult-content resolvers.
-* **Attachment checks.** Executables found by their bytes, also when renamed to `.pdf`; double extensions; right-to-left filename tricks; executables in ZIP files; Office macros; active PDFs; and ClamAV.
+* **Attachment checks.** Executables found by their bytes, also when renamed to `.pdf`; double extensions; right-to-left filename tricks; executables in ZIP files or compressed on their own (`invoice.exe.gz`); text parts, files uuencoded or yEncoded into the text, and other names that mail programs show as attachments; Office macros; active PDFs; and ClamAV.
 * **SPF, DKIM, DMARC and ARC**, DNS blocklists, allow and deny lists.
 * **Language models for close calls.** Ollama, LM Studio, llama.cpp, vLLM, Claude, ChatGPT, Gemini, Mistral, Groq, OpenRouter and any OpenAI-compatible server, on any URL, port and authentication, plus the decision models Cloudflare Clef and TypeSafe Jev. By default the model returns a probability for each verdict in one step instead of writing an answer. Personal data is removed before mail goes to a remote provider.
 * **Every way to plug in.** A Node.js library, a command line, a Postfix and Sendmail milter, a Postfix content filter, a SpamAssassin-compatible spamd server for spamc, Exim and Haraka, an HTTP API and a TCP server.