Anubis ships WebAssembly proof-of-work against AI scraper bots
The firewall guarding LKML, FFmpeg and sourcehut recompiles its browser challenge to bytecode that's harder to spoof — plus, why leaked embeddings aren't safe.
Copy markdown
A year of work, now in bytecode
Anubis, the open-source proof-of-work wall that gates AI crawlers, shipped its long-promised WebAssembly rewrite in the v1.28 pre-release. Author Xe Iaso spent roughly a year moving the browser challenge out of JavaScript and into compiled WASM.
Why WASM beats a JavaScript puzzle
Compiled bytecode is far harder for scraper farms to read, reimplement, and replay than minified JS, and it runs the hash puzzle faster and more uniformly across browsers. The challenge gets tougher to bypass at scale without driving a real headless browser.
Who's actually running this
Anubis already fronts the Linux kernel mailing list archive, FFmpeg, and sourcehut; a Duke pilot clocked it blocking over 4 million unwanted requests a day. If you self-host a git forge, wiki, or docs site getting hammered by training crawlers, this is the update to pull.
Elsewhere: your stored embeddings aren't opaque
The 2025 "Universal Geometry of Embeddings" paper (vec2vec) is back on HN's front page: it translates text embeddings between models with no paired data, then inverts them to recover content. For RAG builders, a leaked vector store can expose what users put in it — even without your original model.