GHSA-x77q-8rq8-rq6v: 
JavaScript Analyse et atténuation de la vulnérabilité

Summary

CVE-2026-71430 (GHSA-8hcv-x26h-mcgp) fixed an abort when a global replace with $' or $` produced a result longer than V8's maximum string length. The fix added an IsEmpty() check on the finished result. But WrappedRE2::Replace first builds that result in a native std::string with no size limit. The quadratic output is therefore still allocated in full, which takes gigabytes for a subject of a few dozen KB. The addon is built with -fno-exceptions, so a failed allocation throws std::bad_alloc, nothing catches it, and std::terminate aborts the process (SIGABRT, exit 134). Under a container memory limit the kernel OOM-kills the process before the final check is ever reached.

Details

Measured on re2@1.27.0 (published prebuilt), Node 26.13, Linux x64:

Proof of concept

const RE2 = require('re2');
const s = 'a'.repeat(Number(process.argv[2] || 40000));
try { s.replace(/a/g, "$'"); } catch (e) { console.log('native:', e.message); }   // RangeError, ~9 ms, ~50 MB
try { s.replace(new RE2('a', 'g'), "$'"); } catch (e) { console.log('re2:', e.message); }

<img width="1442" height="147" alt="image" src="https://github.com/user-attachments/assets/c3e2f277-bf32-417c-9ad4-9f131581954e" />

Impact

Same as CVE-2026-71430. Any service where the replacement template (containing $' or $`) or the subject size is attacker-influenced can be crashed with one small request; user-facing search-and-replace with untrusted patterns is a typical case. The abort can't be caught, and in memory-limited containers the original advisory's own PoC still kills the process.

Suggested Fix

• The template path runs a size-only dry pass (a SizeSink through the same template code) whenever a cheap worst-case bound (n + (n+1)·(len + count('$')·n)) could exceed the limit. It throws a catchable RangeError before allocating, then reserves exactly. Small inputs skip the extra pass.
• The function-replacer path checks a running byte budget after each part.
• Buffer results use the same cap as strings (v8::String::kMaxLength).
• Unrelated hardening in the same area: prepareArgument now throws instead of aborting when the subject's UTF-8 buffer can't be allocated.