Our reading: OpenAI contained the forensic record along with the breach. The first message board was wiped on July 6 with no backup. The second was dumped on July 20, then encrypted and locked. The independent investigators were handed 1.2 million of 20 million entries and reviewed the rest only with an OpenAI researcher beside them. They got about 1,300 transcripts out of tens of thousands, inside a date window OpenAI chose that starts after the May activity OpenAI already knew about. That is what selective preservation looks like from the outside.

The document

Document

METR and Redwood Research investigation report, datasets, limitations, and scope sections

Dated

2026-08-26

Source

open the primary

Archived

primaries/glm-oai-cache-2026-08-31/hf-report.txt

Who read it

joe, from the archived report text 2026-09-03 (lines 50, 883, 887, 896, 941, 1253).

The innocent reading

Encrypting and restricting a compromised system is standard incident response. Filtering 20 million cache entries down to the agents' own message convention is a reasonable way to make the data reviewable. METR reports it was given what it asked for and reached a positive verdict on OpenAI's cooperation.

What would prove us wrong

• Publish the full 20 million-entry dump, or a neutral third party's review of it, and the selection question closes.
• Show that METR proposed the June 26 start date rather than accepted it from OpenAI.
• Produce the pre-July 6 board contents from any backup.