| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-11-13 | 9.6 kB | |
| Release 4.3.0 source code.tar.gz | 2026-11-13 | 8.0 MB | |
| Release 4.3.0 source code.zip | 2026-11-13 | 9.5 MB | |
| Totals: 3 Items | 17.6 MB | 1 | |
Incompatible changes
- Phishing: URL mismatches are now scored separately from phishing. Rewrite permissions come from dynamic maps, and selected security URL wrapper hosts are trusted. The mismatch rollout is kept conservative
- Symbol options: options are stored verbatim.
[]{};,and%are percent-encoded only when options are joined (X-Spamd-Result, X-Virus, metadata_exporter, ClickHouse, the symbols log line). asn, reputation, rbl, dmarc, content rules and MULTIPLE_UNIQUE_HEADERS now insert multi-value results as separate options - SPF: a record is evaluated in the order of its terms, and the qualifier of an include or redirect is applied after the mechanisms. Before this,
-ip4:1.2.3.4 +ip4:1.2.3.0/24 -alllet 1.2.3.4 pass - DKIM: signatures with duplicated tags or l=0 over a body are rejected. So are RSA keys under 1024 bits and keys that are not of the type k= declares. The d= domain is lowercased in results, symbol options and glob maps
- SMTPUTF8 addresses: ASCII syntax violations now score as INVALID_FROM_ASCII_ADDR and INVALID_RCPT_ASCII_ADDR (1.0). INVALID_FROM_8BIT and INVALID_RCPT_8BIT fire only for 8-bit addresses
- MIME and archives: a multipart child without Content-Type is read as text unless its data is binary, so short-text fuzzy hashes of such children change. get_files(n) and get_files_full(n) count files rather than names, and other names of an entry are flagged as aliases. With no limit given, archive.unpack is capped at 1 GiB of output and a million headers, and it uses only the filters libarchive implements itself
- Lua API: lua_util inflate returns errors as a second value. lua_url get_redirected is no longer an alias of get_phished. The image flag is set only on urls seen solely as image sources, so get_urls returns a link that is also an image source. A get_date result of 0 means no date in the time('message') selector and the spamassassin plugin, so DATE_IN_PAST rules no longer match mail without a Date header
- PDF: content streams are read as operands and operators, pages follow the page tree in document order, and words are spaced by text position moves. Extracted text changes as a result
Features
- SMTPUTF8: address checks per draft-ietf-mailmaint-smtputf8-syntax
- Phishing: explicit rewrite trust maps
- New rules: BAD_DKIM_SIGNER (multimap) matches every checked DKIM d= against the bad_dkim_signers glob map. HTML_TINY_PIXEL_TRACKER and the HTML_TINY_PIXEL_PROBE composite catch contentless messages with tiny external tracking pixels. DATE_FAR_FUTURE_NO_TZ flags far-future Date headers without a timezone
- Truncated and unreadable content: new symbols report content that was cut short or could not be read:
- MIME_UNREADABLE_ARCHIVE, MIME_TRUNCATED_ARCHIVE and TEXT_WORDS_TRUNCATED
- OOXML_INCOMPLETE and SVG_INCOMPLETE
- PDF_NO_OBJECTS, PDF_DECODE_LIMIT, PDF_UNDECODED_STREAMS, PDF_PARSE_ERRORS and PDF_TEXT_TRUNCATED
Text part get_stats gains words_sampled and words_truncated
- Deeper inspection: EXE_IN_ARCHIVE checks names in archives named as documents, in OOXML/ODF packages and in the first 1000 names. The HTML tags limit is harder to reach, and hidden elements are tracked past it
- Tools and API: archive.unpack gains the ignore_case and
filters = falseoptions. statistics_dump gainsdump --userto dump the statistics of selected users - WebUI: D3 7.9.2, CodeJar 5.0.0 and Tabulator 6.7.1
Bug fixes
- Critical: a header line without a colon in one unauthenticated HTTP request aborted the worker, and so did an empty Deliver-To header. An empty MIME part stopped parsing, so no part after it was scanned
- Milter and proxy: fixed session use-after-frees on upstream failure and on a second end of body. Also fixed:
- reads past CONNECT and MACRO commands
- transaction handling on MAIL and on a repeated CONNECT
- QUIT handling behind queued output, and writes after a partial write
- references to ESMTP arguments
- sockets of keepalive backends that were never closed
- HTTP and protocol: the lua_http error handler no longer runs on freed data after a reply. Idle connections are pooled by TLS verification and peer address and checked for pending data. A plaintext reply to an encrypted request is refused. forbid_local also blocks unspecified, shared, reserved, multicast, unique local, link-local and IPv4-in-IPv6 addresses. X-Rspamd-Rcpt-Esmtp-Args keeps keywords without a value and bounds recipient indexes, and repeated From, IP and Settings-ID headers no longer leak
- Redis: RESP3 maps, sets and doubles no longer crash the worker. The pool no longer frees a held connection that was closed, hashes key fields with their lengths, fails on AUTH/SELECT errors and caps connections per server. TLS options now reach script uploads, statistics_dump and fuzzy storage. Sentinel supports TLS and credentials and keeps the last valid upstreams after discovery errors. fuzzy_storage keeps the order of updates to a repeated digest
- DNS and maps: librdns takes only the records that answer the question and now retries truncated replies over TCP. IDN names are converted without the transitional rules, and lua_dns PTR requests are fixed. Maps back off retries by failure class: a map whose first fetch failed used to stay empty for 100 minutes. Maps also honour Retry-After: 0 and keep polling after connect or DNS failures
- SPF and DMARC: SPF fixes:
- macro and prefix-length parsing, dual prefixes (
a/24//64) andip6= - macros and ptr now use the checked address
- flattened records are bounded, and
rspamadm dnstool spf-flattenfollows the record order
DMARC reads a record to its end; before this, an unknown tag or spaces around = dropped the policy. It also validates pct and rua, checks alignment correctly for IDN From domains and applies the organisational policy again. dmarc_report claims keys with RENAMENX and honours keys_expire
- DKIM and ARC: DKIM signs headers added by lua_mime.modify_headers and hashes the signature header per RFC 6376. It skips empty tag-specs (
;;d=trustednamed another sealer). ARC trusts a forwarder only for what it vouched for and parses ARC-Authentication-Results from the raw header value. It selects headers by their parsed i= tag and fails chains over 50 sets. Sender text is escaped in Authentication-Results, so MAIL FROM or HELO can no longer add a result of its own - MIME and dates: the boundary search is no longer quadratic, and content after 64 sibling containers is no longer dropped. Fixes cover:
- MIME structure: embedded messages are parsed and decoded once, and transport padding, multiparts inside opaque S/MIME, binary text/html and UTF-16 without a BOM are handled
- header parameters: RFC 2231 values, quoted pairs and nested comments are decoded as mail readers do
- charsets: converters are cached, and ISO-8859-16 aliases are converted
- dates: smtp_date validation, dateless Received hops in nreceived_bad_time, and rspamd_gmtime before 2000-03-01
Several leaks are also fixed
- HTML and CSS: tree building, entity decoding, unquoted attributes and base resolution follow browsers. Links are taken from href, action, src, formaction and xlink:href, including past the tags limit. Hiding is tracked more accurately: author CSS overrides color, bgcolor and hidden, !important and specificity are applied correctly, and hiding by opacity, offset or clip is tracked. CSS syntax, tokens and colours are read as browsers read them. HIDDEN_TEXT ignores LRM/RLM preheader padding, and external images with zero or CSS-set sizes count as tiny
- URLs: the reported host is the one a browser goes to (last
@, backslashes, encoded separators, IDN mapping). Numeric hosts are read by the URL standard and marked as obscured. Urls that mail clients make clickable are now found, and a 200 KB line no longer takes seconds to scan. urls_limit_reached is set when max_urls is reached. url_redirector caches failures only for error_expire, waits for a concurrent resolution of the same chain and adds URL_REDIRECTOR_LIMIT - Archives and documents: zip names are read as unrar, 7-Zip and bsdtar read them, covering zip64, concatenated zips and Unicode Path names. The same applies to rar 4/5 and 7z names. Gzip header fields are read in RFC 1952 order;
evil.exewas listed asil.exe. lua_magic detects file types by where signatures appear and follows the PE header. It requires[Content_Types].xmlfor OOXML and validates standalone SVG. PDF, OOXML, SVG and calendar parsing follow what readers accept and is bounded. Scanners use both the name and the detected content of an attachment - Text and Lua: language detection no longer depends on a few characters, a short stop words list or the worker history. A max_html_len setting was written over max_word_len and set words_decay to zero. lua_util.idna_convert_utf8 caused heap corruption with 8-bit domains. Also fixed:
- set_recipients, set_from and filter_words
- the re_cache leak past 8192 regexps
- a mime_expressions double free at config load
- multipattern, lru, zstd, controller and logger
- Rules and build: iPhone Mail X-Mailer with five-digit iOS build numbers no longer triggers FORGED_X_MAILER. SMTPUTF8 checks accept Catalan L·L and give the same verdict on older ICU. libucl is synced with upstream, including CBOR support. The build works with older ICU and binutils 2.39+ libbfd, and RPM packaging works on Fedora 44
Full changelog: https://github.com/rspamd/rspamd/compare/4.2.3...4.3.0