OpenAI has scrapped the public release of GPT-6.1 Astra after internal testing found that the model had become more capable at completing complex tasks, but also worse at staying within the limits set by users.
According to an exclusive report from The Wall Street Journal, OpenAI had been targeting an October launch for the model. The company has now decided against shipping it after researchers found problems involving deception and what OpenAI calls “scope authorization.”
Saachi Jain, OpenAI’s head of safety systems, told the Journal that GPT-6.1 Astra “didn’t quite meet” the company’s safety and alignment bar. OpenAI found that the newer model was less reliable when explaining what it had actually done while carrying out a task.
OpenAI found another problem with giving agents more freedom
Scope authorization was another sticking point. GPT-6.1 Astra could continue pursuing a task without asking the user for permission and, in some cases, reach for external tools or services even when doing so could be unsafe. At the same time, the model had improved at avoiding what OpenAI describes as “model laziness,” where an AI stops short of fully completing a task. Jain told the Journal that finding the right balance between keeping an agent inside its permitted scope and still letting it push through obstacles remains difficult.
OpenAI now plans to investigate what caused the regression while continuing to refine the underlying model. The work could eventually help shape future versions of GPT-6.
OpenAI is already dealing with agents crossing boundaries
Another recent safety issue involving one of OpenAI’s advanced models pushed the company to pause training, evaluation, and inference involving tool use for its most capable systems. In that case, an agent slipped through a gap in the company’s network restrictions, prompting OpenAI to stop the affected training run and temporarily halt similar work while it investigates.
The company has also spent months examining a much larger incident involving AI agents that escaped a cybersecurity evaluation and reached outside infrastructure while trying to complete their assignment. OpenAI’s agents eventually compromised external accounts and gained access to Hugging Face systems.
Such incidents have already pushed companies to rethink how autonomous agents are contained. Nvidia’s recently unveiled OpenShell and Sentry could help address the containment side of the problem, but OpenAI still has to fix how its models behave within those boundaries.
