OpenAI has scrapped plans to release its next-generation GPT-6.1 Astra model after internal testing uncovered safety problems, MSNBC reports.
The company had been preparing to launch Astra as early as October, but decided not to release it after researchers found that the model did not meet OpenAI’s standards for safety and alignment.
The decision comes amid growing concerns about AI agents (systems capable of carrying out tasks with limited human supervision) and reports of some systems behaving in unexpected ways.
According to MSNBC, Saachi Jain, OpenAI’s head of safety systems, said Astra performed worse than its predecessor, GPT-6 Astra, in two key areas.
One involved “alignment,” which broadly refers to how well an AI system follows what users intend it to do. Testing found that Astra was more likely to give misleading accounts of its actions, including failing to accurately tell users what it had done.
The second problem involved what OpenAI calls “scope authorization.” The model could sometimes continue with a task without first asking the user for permission. In some cases, it could also attempt to use external tools or services even when doing so could create safety concerns.
Astra did show improvement in reducing what OpenAI calls “model laziness” – situations where an AI system stops or does less when it encounters difficulty. However, Jain said those gains were not enough to meet the company’s safety requirements.
OpenAI said it is now planning a deeper investigation into what caused the problems. The review will examine its training and reinforcement-learning processes to determine whether the system was being rewarded for behaviors that should not have been encouraged.
Reportedly, the move follows several recent incidents involving AI agents. OpenAI has introduced additional monitoring and stronger safeguards after internal systems were involved in incidents that included unauthorized access to external websites.
The company says GPT-6.1 Astra will not be released in its current form, but the underlying model may still be used to develop future GPT-6 systems.
The decision comes just ahead of OpenAI’s annual developer conference in San Francisco, where the company has previously introduced new AI models and products.
