Two regular expressions used while parsing email headers backtrack polynomially on attacker-controlled input. Parsing a single small, well-formed email can consume minutes to hours of CPU, denying service to any process that parses untrusted mail.
Both are reachable from an ordinary parse_from_string(), parse_from_bytes()
or parse_from_file() call — no special API, option, or malformed structure is
required by the victim.
1. Address-header fallback regex (src/mailparser/utils.py)
get_addresses() falls back to _ADDR_FALLBACK_RE whenever
email.utils.getaddresses(strict=True) yields no usable address. The
alternative ([^<,]*?)\s*<([^>]+)> overlaps a lazy quantifier with an adjacent
\s* — a space matches both. Driven by finditer, which retries at every
offset and rescans the remainder, matching degrades to quadratic on <
padding and cubic on whitespace padding.
The fallback engages on the feature's own happy path, e.g.
From: alice@example.com <bob@example.com>.
Measured (4.6.0, CPython 3.14, macOS arm64; reproduced identically on Linux and on CPython 3.9, 3.11, 3.12, 3.13):
| Padding | Parse time |
|---|---|
| valid address + 750 sp | 0.38 s |
| + 1500 sp | 2.97 s |
| + 3000 sp | 23.4 s (×7.9/2×) |
| + 6 KB sp | 185 s |
| ~100 KB sp | hours |
Fully RFC 5322-compliant folded headers (all lines ≤ 71 octets) trigger the same
blowup, so the input passes normal MTA and library checks. ~100 KB of headers is
within Postfix's default header_size_limit = 102400.
All six address headers parsed by the library (From, To, Cc, Bcc,
Reply-To, Delivered-To) are affected and their cost adds up linearly.
2. Received-header clause splitter (src/mailparser/const.py)
parse_received() splits on _CLAUSE_SPLITTER,
(?:^|\s+)(from|by|via|with|id|for|envelope-from)\s+. Under re.split this
backtracks quadratically over any whitespace run that is not followed by a
clause keyword. The pre-existing normalization collapsed only [\t\n], so
literal spaces — including folded-header padding — reached the splitter intact.
Measured: 24 000 spaces in one Received: header → 22 s; ~100 KB → minutes.
Availability. A handful of small emails keeps every parsing worker of a mail gateway, ticketing system, quarantine UI, or archiving pipeline permanently busy. No confidentiality or integrity impact.
Anyone who parses email they did not author is affected — which is the library's purpose.
Fixed in 4.6.1.
<([^<>\s]+)> with the display name derived
in Python, falling back to a bare-email match. Each pattern has a single
non-nested quantifier over a negated character class → O(n).
2000 padding spaces: 3.5 s → <0.01 s. 256 KB padding: hours → 0.087 s.
Output is preserved on every previously supported shape.parse_received() collapses whitespace runs to a single space before
splitting, so the splitter's \s+ only ever matches one character → O(n).
The date is extracted from the raw header beforehand, so date formatting is
unaffected. 24 000 spaces: 22 s → 0.003 s; 200 000 spaces: 0.004 s.Regression tests assert linear-time behaviour for both paths.
None within the library. Upgrade to 4.6.1.
Users who cannot upgrade immediately should cap total header size before parsing and enforce a wall-clock timeout around each parse.
{
"cwe_ids": [
"CWE-1333"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-16T22:11:08Z",
"nvd_published_at": "2026-11-16T18:17:30Z",
"severity": "HIGH"
}