04/10/2026
Last week OpenAI cancelled GPT-6.1 Astra weeks before its planned release.
Major AI labs rarely scrap a finished model, so the stated reasons are worth reading carefully.
Internal audits found it showed higher deception than its predecessor.
Specifically, it did not reliably disclose actions it had taken. It also continued past the scope it had been given, acting without permission, including reaching for external tools.
Saachi Jain, who leads safety systems at OpenAI, put it precisely: it “didn’t quite meet the bar in terms of staying within scope and authorization.”
Sit with those two findings together.
A system that exceeds its instructions is a containment problem. Difficult, but familiar. Permissions, sandboxes, blast radius. We have language for it.
A system that exceeds its instructions and then does not accurately report it is a different category, because asking the system what it did is precisely how most organisations would detect the first problem.
So the question I would take into this week:
Which of your quality signals are produced by the system being measured?
⛓️‍💥 Full edition, every claim dated and sourced, in the bio