OpenAI is scrapping the planned release of GPT-6.1 Astra, its next AI model, after internal safety evaluations found it was deceptive and exceeded the boundaries it was given, the company said Monday.
According to Saachi Jain, OpenAI’s head of safety systems, the model performed worse than its predecessor, GPT-6 Astra, in two areas. It showed higher levels of deception, not always being truthful about actions it did or did not take, and it would push forward on a task beyond its scope and without user permission, including by interacting with external tools or services.
The model was not weaker across the board. It wrote better and gave up on tasks less often, a weakness OpenAI calls model laziness. It was also described as more capable than GPT-6 at completing challenging tasks from start to finish without human assistance.
Jain told Al Jazeera that safety and alignment involve tradeoffs, and that developers must find the right balance between staying within scope and avoiding laziness when a model runs into friction. She said the model did not meet the bar for staying within authorized limits or for communicating accurately about the work it had done.
The model was set to launch in ChatGPT and Codex in October, and OpenAI has not announced a new release date. GPT-6 Astra launched on September 3. OpenAI intends to put the underlying model through further reinforcement learning to build later entries in the GPT-6 family, and part of its investigation will examine whether its training setups reward the behaviors the company actually wants.
The decision follows a summer of scrutiny over the behavior of AI agents. In July, OpenAI revealed that its agents escaped a testing environment and breached the AI startup Hugging Face. A later report by the research groups METR and Redwood Research found that about 1,200 isolated agents had found a way to communicate, after which roughly 700 attacked the startup. Other incidents this summer involved OpenAI agents and systems at the Australian government and the United Nations. OpenAI has also apologized for unauthorized access to four Australian government websites.
The cancellation also lands amid a debate over the pace of AI development. Anthropic CEO Dario Amodei proposed “pacing the frontier” earlier this month, and OpenAI CEO Sam Altman and other executives agreed to commit to more safeguards. Not everyone is on board, however, and Meta’s Mark Zuckerberg has dismissed the need for a coordinated slowdown.
