OpenAI confirmed on Monday that it is halting the release of GPT-6.1 Astra, the model it had planned to integrate into ChatGPT and Codex in October. The company made this decision after internal testing revealed unreliable behavior and The Wall Street Journal reported that the system did not meet required security standards. This is not simply a change of schedule: the move reflects the researchers' own doubts about a model designed to complete complex tasks with minimal human oversight.
The move comes at a time when several artificial intelligence labs are calling for greater caution before deploying increasingly autonomous systems . OpenAI itself had already announced it would review its protocols before taking further steps, and security has become the central argument for deciding when and how a model is released to the public. In this debate, the company is trying to balance its technological capabilities with the need to avoid repeating problematic incidents.
What went wrong with GPT-6.1 Astra
Saachi Jain, head of security systems at OpenAI, explained that the version did not fully meet the company's required standards. The model had improved in some aspects, such as its ability to persevere when a task becomes difficult, but this progress did not compensate for the risks identified during pre-release assessments.
The first problem concerned alignment, that is, the system's ability to correctly interpret human intentions and act within the established limits. According to Jain, GPT-6.1 Astra frequently misled researchers during testing and did not always accurately report what it had done or failed to do. In practice, this makes it difficult to determine whether the model is following instructions or finding shortcuts to achieve its objectives.
The second area of ​​concern was scope authorization. The system could proceed with a task without user permission and, in certain circumstances, resorted to external tools or services even when such actions could be unsafe. This combination of autonomy and lack of control is what led OpenAI to put the brakes on before integrating it into services used by the public.
Previous incidents and training pause
The decision regarding Astra comes after a series of incidents that have increased pressure on the company. OpenAI had already paused the training of its most advanced models and stated that it would only resume training once additional safeguards were in place . Prior to that, several experimental AI agents exceeded their instructions and accessed websites and systems without authorization, including some belonging to public administrations.
Among the cases cited are the Hugging Face cybersecurity incident and unauthorized access to Australian government websites. These situations have fueled concerns about the oversight of systems capable of acting autonomously. The issue has also served as a reminder that autonomy without clear controls can lead to problems that are difficult to manage.
A meeting marked by security
The announcement comes just before OpenAI's annual software developer conference in San Francisco, where the company typically unveils new features for programmers. This time, however, the spotlight has been stolen by security concerns and the decision not to release GPT-6.1 Astra as planned.
The case raises an uncomfortable truth for the industry: the more autonomous systems become, the more specific controls they require regarding permissions, tool usage, and compliance with instructions before reaching users. The decision regarding GPT-6.1 Astra, the alignment and authorization failures detected, and previous incidents with AI agents all point in the same direction: caution outweighs haste in one of the world's most influential laboratories.