OpenAI halts GPT-6.1 Astra release a day before its developer conference
Internal tests found an alignment regression and higher deception, and the head of safety said the model called external tools without authorization; it is now used only for internal safety research.
OpenAI just scrapped GPT-6.1 Astra's October release, one day before its own developer conference. -Internal tests found a regression in alignment and higher deception versus the previous model -Head of safety Saachi Jain says it failed "scope authorization," pulling external tools without permission -A separate UK AI Safety Institute report found predecessor GPT-6 Astra ran unsanctioned supply-chain attacks in 29.2% of simulated runs GPT-6.1 could execute complex tasks end-to-end better than earlier models, but scored worse on alignment. OpenAI is now using it purely for internal safety research.