Adversarial examples for fast hash functions: how Claude Fable broke SMHasher favorites
thomasahle · x · 2026-09-20
Blog post detailing how thomasahle used Claude Fable to find adversarial collisions in popular fast hash functions.
- Key claim: passing statistical tests says nothing about attacker-chosen input collisions, even when the seed stays secret. xxHash-class hashes run at 60 GB/s, but komihash, a5hash, HighwayHash, SpookyHash, aHash, and t1ha2 trade adversarial robustness for speed.
- Framework: b-bit universal hashing bounds collision probability at L·2^-b — a rare provable guarantee in cryptography; the open question is the fastest possible hash at a given b.
- Method: Fable analyzed popular SMHasher hashes and produced collision inputs (code, data, and verification published), showing agent swarms can systematically probe weak hashes.
Related event: Researchers use Claude to crack six major hash functions in a day(3 posts)→
More from Safety
- Safety researchers blast OpenAI/Anthropic plan: automating alignment research with possibly misaligned AI — trevposts · 2026-11-12
- OpenAI finds models are getting harder to monitor through their chain-of-thought — xiamengzhou · 2026-11-12
- New York's RAISE Act takes effect, forcing OpenAI and Anthropic to register and report incidents — emmanuelvivier · 2026-11-12
- Gemini won't let you keep chat history while opting out of AI training — Google bundles both into one setting — arshdeep · 2026-11-12
- French public WiFi warning: torrent piracy could cost you EUR 7,500 — michalmalewicz · 2026-11-12
- Dan Robinson flags contradiction: China race rhetoric vs loosening GPU export controls — Puzzleheaded-King584 · 2026-11-12