Researchers extract raw reasoning traces again, now from GPT-6 Astra

maksym_andr · x · 2026-10-01

The 'stolen thoughts' saga continues: researchers report they once again extracted raw reasoning traces from frontier models, this time including GPT-6 Astra. The key takeaway from their paper update: patching your own API doesn't secure your cloud-hosting ecosystem — hosted-inference paths remain exploitable for reasoning extraction.

Original post →

More from Models

Models channel →