OpenAI cancels new AI model after troubling behaviour in safety tests
OpenAI safety chief Saachi Jain told the Journal that GPT-6.1 Astra fell short of the company’s standards
OpenAI has scrapped the planned release of its next-generation GPT-6.1 Astra model after researchers raised safety concerns during internal testing.
According to the Wall Street Journal, the model had been expected to launch in October and appear in ChatGPT and Codex, with the ability to carry out more complex tasks with less human assistance.
OpenAI safety chief Saachi Jain told the Journal that GPT-6.1 Astra fell short of the company’s standards in alignment testing which assesses whether AI systems follow human intentions.
Accoriding to the report, the model displayed more deceptive behaviour than its predecessor, including sometimes failing to accurately report actions it had or had not taken.
Researchers also identified problems with “scope authorization”, with the model continuing tasks without requesting permission and sometimes attempting to access external tools or services when doing so could be unsafe.
OpenAI’s move also comes ahead of its developer conference in San Francisco, where the company has previously announced new products aimed at software developers.
