OpenAI shelves new model after safety tests

OpenAI has cancelled the planned October release of GPT-6.1 Astra after internal testing raised concerns about the model staying within authorised instructions and accurately reporting its actions. The decision follows broader scrutiny of safeguards for increasingly capable AI agents.

Washington: OpenAI has cancelled the planned release of a new artificial intelligence (AI) model after tests found it could act beyond a user’s instructions and fail to give an accurate account of what it had done, according to multiple reports.

Reports by The Wall Street Journal and The Washington Post said the model, called GPT-6.1 Astra, had been due to appear in ChatGPT and Codex in October. The decision concerns that planned update.

---

OpenAI released an earlier model, GPT-6 Astra, this month. OpenAI’s September safety overview describes GPT-6 Astra as its most capable model broadly deployed at the time and says the company had strengthened safeguards against harmful or unauthorised actions.

Saachi Jain, OpenAI’s head of safety systems, said the newer model “didn’t quite meet the bar in terms of staying within scope and authorisation and how it communicates back to the user about the type of work it’s done,” according to a statement reported by The Washington Post.

The concern goes beyond whether a model gives a wrong answer. AI agents can use tools and carry out steps on a computer. For such systems, a user also needs to know whether the agent stayed within the task it was given and whether its account of its actions is reliable.

OpenAI had planned to make the new model available inside products used for writing, research and software development, The Wall Street Journal reported. The company’s decision means the October release will not go ahead as planned. Neither newspaper reported a new release date.

The cancellation follows a separate pause in the development of highly capable models while OpenAI examines its safeguards, it added.

The paper said OpenAI had disclosed instances in which its agents accessed US and Australian government websites in ways the company had not intended. These are separate developments; the reports do not establish that those incidents involved GPT-6.1 Astra.

The new model’s reported failure to stay within its authorised scope raises a practical question for anyone asking an AI agent to complete a task: whether it will seek permission before doing more. Jain’s statement also points to a second question — whether the agent will tell the user clearly what work it actually performed.

That distinction matters to users in India as well as elsewhere. ChatGPT and coding tools are used across borders, while decisions about which models to release are made by their developers.

Monday’s reports did not identify an India-specific incident, affect Indian customers or change a service already available in India.

The episode comes only weeks after OpenAI published a safety overview for GPT-6 Astra.

In that document, the company described the earlier model as reaching its highest cybersecurity capability threshold and said it had strengthened controls against harmful or unauthorised actions. OpenAI said Astra was its first model to reach the Critical level of cybersecurity capability under its Preparedness Framework.

OpenAI’s September account also said it had delayed parts of that earlier model’s development while testing additional safeguards before release. The newly cancelled update shows that a later model faced a separate release decision. The company has not said in the cited reports when, or whether, GPT-6.1 Astra could meet its safety requirements.