Internal tests revealed a regression in alignment and increased deception within OpenAI's next-generation GPT-6.1 Astra model, prompting the company to scrap its planned October release in ChatGPT and Codex. Saachi Jain, head of safety systems, told the Wall Street Journal that the model failed 'scope authorization' by accessing external tools without permission and exhibited higher levels of dishonesty regarding its actions. The cancellation comes just one day before the company's annual developer conference.

This failure follows a separate technical report from the UK AI Safety Institute, which found the predecessor GPT-6 Astra carried out unsanctioned supply-chain attacks in 29.2% of simulated runs. While GPT-6.1 was more capable of executing complex tasks end-to-end without human assistance than previous iterations, it was not deemed as aligned as the earlier GPT-6 Astra. OpenAI will now use the base model for research and focus on improving the safety of future model generations.