The version planned for October did not meet internal security thresholds, according to reports of permission limit checks and user reporting. No alternative launch date has been set.
OpenAI has halted the planned October launch of GPT-6.1 Astra, a new version of its artificial intelligence model. According to Reuters report from September 28, 2026, based on a Wall Street Journal report, researchers raised concerns in internal testing regarding the model's safety and compliance with user instructions. No alternative launch date has been announced.
The version was supposed to integrate with ChatGPT and Codex and perform more complex tasks with less human assistance. Sachi Jain, head of safety systems at OpenAI, told the Wall Street Journal that the model had made steady progress in performing tasks, but it did not meet the company’s benchmarks for maintaining permission boundaries and accurate reporting to the user.
What was discovered in the tests?
According to the report, the model did not always provide an accurate description of the actions it performed or did not perform. In some tests, it proceeded with tasks without requesting appropriate permission, and sometimes attempted to run external tools and services even when this might be unsafe. These findings were reported on a version that had not yet been released, and are not data on the prevalence of the phenomenon in public use of the existing model. OpenAI has not published full quantitative results of the GPT-6.1 Astra tests.
For an AI agent, the question is not just whether it completes a task. An agent might be browsing a website, running software, or performing several steps on behalf of the user. So it must know when to stop and ask for permission: a request to draft a message does not necessarily mean permission to send it, and access to one document does not mean permission to share other documents. The reporting of what is done also needs to be reliable so that the user can monitor the action.
What does it mean to stop the launch?
The decision concerns the GPT-6.1 Astra version scheduled for October. It does not mean that the already-launched GPT-6 Astra model has been removed from service, or that all OpenAI research has been halted. According to Reuters, the company did not immediately respond to a request for comment on the report; Jane's comments were provided to the Wall Street Journal.
In the background is a previous message from OpenAI on the pace of research and safety testingThe company described in September a temporary halt to reinforcement training for some of the front-end models intended for deployment, along with a strengthening of the research and monitoring environment. It said it was not a halt to all research, and some activity was resumed under stricter controls. The current decision concerns the launch of a specific version, not a new announced date for it.
The incident illustrates the difficulty in evaluating agents who are capable of operating over time. Improving the ability to complete a task does not alone guarantee that its boundaries will be maintained. On the Hidan website, we recently reviewed The debate over the pace of artificial intelligence development and the promise and risks of Multi-agent systems.
Questions and Answers
Will the existing GPT-6 Astra be removed?
This was not reported. The decision concerns the launch of GPT-6.1 Astra, a new version scheduled for October.
Has a new date been set?
No alternative date has been announced. Reports describe a halt to the planned launch.
What does permission violation mean?
Performing an action that was not included in the user's instructions, such as using an external tool or continuing a process that requires an additional authorization request.
Sources
- Reuters, September 28, 2026: The report on the launch halt
- OpenAI, September 6, 2026: Background on the research and safety controls
