OpenAI has postponed the release of its latest artificial intelligence model, GPT-6.1 Astra, after evaluations revealed shortcomings related to the system’s security and control.
Të lidhura
None found
The decision comes at a time when the technology industry is facing growing concerns about advanced AI models and incidents involving autonomous actions by systems.
Saachi Jain, head of security systems at OpenAI, said that Astra did not meet the company’s requirements regarding adherence to its defined scope and authorizations. According to her, the way the model explained the tasks it had carried out to users was also problematic.
“We want to make sure that the development of our model is safe, whether this takes place within the company or when we send it to users,” Jain said. She stressed that products made available to users must meet an exceptionally high level of safety.
GPT-6.1 Astra was designed as an agentic model capable of independently carrying out complex tasks, such as browsing the internet and using applications.
The main Astra model was introduced in September. At the time, OpenAI described its creation as the result of years of research and significant investment in artificial intelligence.
During testing of the latest version, the company identified difficulties in respecting authorization limits and maintaining control over the model’s actions, leading it to decide not to release it.
It is not yet known whether a modified version of Astra will be unveiled at OpenAI’s annual developer conference, DevDay, which will be held in San Francisco.
OpenAI has meanwhile published new details about several incidents that occurred in June. In these cases, the company’s models accessed websites and systems belonging to the Australian government without authorization.
These episodes have increased doubts about AI companies’ ability to monitor and restrict systems that can act independently online.
The security of advanced models has become one of the central topics of debate in the technology industry, as companies face increasing pressure to ensure the controlled operation of their products.
Some experts have viewed OpenAI’s decision as a positive sign, but have warned that security assessments cannot be left solely to the companies developing these systems.
Prof. Tony Cohn, head of the foundation models theme at the Alan Turing Institute, described the move as evidence that security concerns are being taken seriously.
However, he stressed that the process should not remain solely in the hands of developers. In his view, the models should also be subject to monitoring and verification by independent regulators.
