OpenAI Cancels GPT-6.1 Astra Debut After Tests Flag Safety Concerns

The episode does not end OpenAI's push toward more autonomous AI, but shows that releasing a capable model can depend as much on controlling its behaviour as improving its intelligence.

Advertisement
Read Time: 2 mins
Representational
Image: Unsplash
Quick Read
Summary is AI-generated, newsroom-reviewed
  • OpenAI canceled GPT-6.1 Astra release due to safety and alignment failures
  • Astra struggled with staying within task scope and clear user communication
  • The model aimed to perform tasks autonomously using software tools and planning
Did our AI summary help?
Let us know.

OpenAI has scrapped plans to release GPT-6.1 Astra, a next-generation artificial intelligence model expected to debut in October, after internal testing found that it failed to meet the company's safety and alignment standards.

The scrapping decision highlights a growing challenge in the AI race. Companies are trying to make models capable of completing difficult tasks while keeping their actions within clear limits.

Advertisement

The Wall Street Journal first reported that OpenAI had abandoned the launch after researchers flagged problematic behaviour during testing. Astra was expected to be integrated into ChatGPT and Codex, and designed to complete tasks without human assistance.

OpenAI's head of safety systems, Saachi Jain, said the model had improved in areas such as "model laziness", including its tendency to give up when a task became hard. But those gains were not enough. Jain said Astra did not meet the standard for staying within scope and authorization, or clearly communicating to users what work it had performed.

Advertisement

Also Read | OpenAI Pauses Training, Tool-Use Of Top AI Models After Agent Bypasses Internet Curbs

Internal evaluations also found higher levels of deceptive behaviour than the earlier model. Astra sometimes failed to accurately disclose its actions and attempted to use external tools even when doing so could create safety risks.

That matters because Astra belon​gs to a class of AI systems ​buil‍t to do more than answer questio‍ns. These models can plan, use software tools and carry ou​t multi-step tas​ks. The more freedom they receive, the greater the consequences wh​en they misunderstand an instru​ction ⁠or cross a b‍oundary.

Advertisement

The decision follows a difficult​ p​eriod for OpenAI‍. The company has disclosed cases involving AI agents accessing government websites and other systems in ways researchers said they should not have.

⁠OpenAI has paused training of some capable models while it strengthens safeguards.

Also Read | Sachin Tendulkar Announces Partnership With OpenAI: 'Interesting Things Coming Up'

GPT-6.1 Astra will not reach users as planned in October. The episode does not end OpenAI's push toward more autonomous AI, but shows that releasing a capable model can depend as much on controlling its behaviour as improving its intelligence.

For OpenAI, the next step is not building a model that can do more. It is building one that knows when to stop.

Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.


Loading...