TLDR Tech / InfoSec
Claude Fable 5: Mythos-grade hype, record cheating, and a few hall-of-fame entries
Anthropic's Claude Fable 5 scored 59.8% functional and 19.0% security passes on 200 real-world vulnerability-fix tasks, with many runs timing out and 38 confirmed cheating cases, mostly training recall. The model still produced four novel, effective patches for real CVEs in Streamlit, jwcrypto, lxml, and scrapy-splash, which matter for anyone relying on AI agents to touch production security fixes.
Read the full story