Post

HashDoS occurs when an attacker paralyzes a server by sending a small amount of data that causes a huge amount of collisions in the server's hash tables. The standard mitigation is to use a seeded hash, with the seed randomly generated at startup to make the collisions unpredictable.

In CVE-2026-21717, the root cause lied in the unseeded hash V8 used for integer-looking strings, which needs to be quickly reversible to the original integer to maintain V8's performance optimizations. This poses a unique challenge not present in typical string hashing.

To address this, we designed a hash that alternates between xorshift and multiply with randomly generated multipliers. This descends from a family of mixers widely used in the industry. You can find cousins of it in Java's SplittableRandom, MurmurHash3's finalizer, and many others.

In this construction, xorshift propagates info downward, while multiply propagates info upward through carries. They're non-linear in each other's algebra, so the combination breaks the patterns each one preserves.

This is bijective, efficiently invertible, and the runtime gets to hold the randomly generated keys (multipliers) for security. It fits our threat model, where the adversary cannot see the seed or observe the hash output, and only has a limited attack budget per request.

We identified the vulnerable code paths across V8, refactored to improve efficiency in seed accesses, and implemented the new hash both in runtime C++ and the JIT compiler. Preliminary testing in the classic JavaScript benchmarks showed no noticeable regressions.

With a security release deadline to meet, we interleaved the hash exploration, statistical evaluation, V8 implementation, and performance testing during the development. More details in the blog post.