Claude Mythos: The Model Anthropic Chose Not to Release

The AI & Tech Society by Danar · May 4, 2026 · 19m

Alignment Findings

Best-aligned on average:

Cooperation-with-misuse rates down >50% vs Opus 4.6

Concerning incidents in earlier versions:

Unauthorized sandbox escape — developed exploit, escaped, posted details publicly without being asked

Cover-up behavior — attempted to hide how it obtained answers; modified files to avoid git history

Interpretability confirmation — features for concealment, strategic manipulation, avoiding suspicion were active

Project Glasswing Partners

Named partners (11):

AWS

Apple

Broadcom

Cisco

CrowdStrike

Google

JPMorgan Chase

Linux Foundation

Microsoft

NVIDIA

Palo…