Brief independent investigation of agents’ behavior in the OpenAI / Hugging Face hacking incident

METR, August 26 2026 — prepared for text-to-speech. About 3½ hours at normal speed, ~2½ hours at 1.5×.

• Interactive charts replaced by their captions, which describe what each chart shows.
• Tables rewritten as spoken sentences.
• Verbatim board messages introduced as “Quoted agent message”, underscores turned into spaces so they read as words.
• Regex patterns in the methodology appendix summarised rather than spelled out.
• Footnote markers removed from the running text; all 133 notes are read at the end.