Anthropic used Claude Mythos for 60 hours and found crypto flaws at a $100K cost
vista8 · x · 2026-07-29
Anthropic researchers asked Claude Mythos to hunt for mathematical flaws in cryptographic algorithms.
- The run lasted 60 hours and cost about $100,000 in API usage.
- It reportedly found theoretical weaknesses in both HAWK and a simplified AES variant.
- The more interesting part was the workflow: the model often seemed to think the problem was impossible, wanted to give up, or looked for shortcuts.
- The researchers’ job was less about feeding data and more about pushing, correcting, and motivating the model to keep digging.
The post argues that prompt engineering and human guidance are undervalued in research settings, because models can also “feel stuck,” get lazy, and need constant steering toward publishable results.
Related event: Anthropic's Claude Finds Weaknesses in HAWK and Reduced-Round AES(39 posts)→
More from Models
- Claude Code adds idle compaction to save 1h prompt cache on idle sessions — EricBuess · 2026-10-16
- Thrive Holdings launches real-work AI evals from 75+ businesses, benchmarking 13 models — samaysham · 2026-10-16
- Microsoft's MAI-Transcribe now handles 61% of all transcription on OpenRouter — mustafasuleyman · 2026-10-16
- Reflection AI Releases Open-Weight 501B MoE Beam for Agentic Coding on AWS — AWS ML Blog · 2026-10-16
- Dev RL-tunes Qwen 27B against slop with Meta's XAR method, lands top-4 on EQbench — Mysterious_Hearing14 · 2026-10-16
- HotDogBench tests 10 LLMs on 1,000 hot dog photos — all within each other's error bars — josh_wills · 2026-10-16